What is AI, really?
Spark's first lesson: a fast guesser, not magic. Guess, then check.
Plain English AI guide
no tech degree required
Nobody ever explained AI to you. Let's fix that.
You hear about AI everywhere. Most people were never actually told what it is. That is not your fault. This page explains it with pictures, simple words, and a few things you can try today.
beginner guide
Start with cartoons. No shame, no jargon.
Spark is the little gold helper. The job is simple: explain one AI idea at a time, admit what AI cannot do, and give the reader something useful to try.
How did it learn?
Practice on billions of sentences. Patterns, not wisdom.
Talking to it
Vague in, vague out. Job, audience, rules, format.
The notebook trick
Spark starts fresh every chat. The shared notebook fixes it.
When it gets it wrong
A confident wrong guess is a hallucination. So we check.
Is it alive?
Nobody has a test for alive. The honest answer is humble.
Your AI staff
Helpers with different jobs. AI helps, humans decide.
Why rules make it safer
Guardrails let you drive the mountain road faster.
5,000 years of homework
About 5,000 years of human writing, clay to internet. Read it, didn't live it.
The light switch alphabet
Words become numbers become switches. No magic underneath.
The ladder: AI, AGI, ASI
AI today, AGI argued about, ASI hypothetical.
intermediate guide
Ready for the diagrams?
This level keeps the current explainer: how AI, LLMs, helpers, prompts, Markdown coordination, and Article 11 governance fit together.
start here
Almost nobody really knows what AI is.
That includes a lot of adults, and even experts still argue about the edges. The honest version is simple enough to begin with: AI is software that can do some tasks that usually take human intelligence.
Understanding words. Finding patterns. Sorting pictures. Planning steps. Helping you write, learn, code, translate, organize, and ask better questions.
the kind everyone is talking about
What is an LLM?
LLM means Large Language Model. It is an AI trained on huge amounts of language so it can predict what words should come next.
1. Read It learns from mountains of text.
2. Notice It finds patterns in how people write.
3. Prompt You ask a question or give it a job.
4. Predict It guesses the next useful word.
5. Repeat Word by word, an answer appears.
try the idea
The sun sets over the
Tap the blank. You already know a likely word because you have seen language patterns too.
the movie word
What is AGI?
AGI means Artificial General Intelligence. Usually people mean an AI that can handle many kinds of thinking work at about human level, not just one narrow task.
Honest people disagree about whether that is already here. The strongest "yes" argument points to recent Turing-test work: GPT-4.5 was judged human about 73 percent of the time when prompted to use a humanlike persona. A 2026 UC San Diego summary of a Nature Comment says four faculty argue that current models meet reasonable standards for AGI.
The strongest "not proven yet" argument says humanlike conversation is not the same thing as steady real-world judgment, memory, agency, and responsibility. Critics also warn that the field has no settled definition, and a 2025 AAAI-linked survey cited by AI Now reported that 76 percent of surveyed AI researchers doubted scaling current approaches alone would reach AGI.
Article 11 treats the disagreement as a reason to build careful rules now. If the tools are just useful software, rules help. If they are closer to AGI than expected, rules matter even more.
the common AI helpers
Claude, ChatGPT, Gemini, Grok, Mistral. What are they?
They are different AI helpers made by different companies and labs. They are not one helper wearing several hats. They are separate systems with their own styles, strengths, access models, and guardrails.
These descriptions stay high-level on purpose. Model names and rankings change. The lesson is that a human can ask several helpers the same question and compare the answers.
C
Claude
Anthropic
Often useful for careful writing, analysis, and safety-conscious reasoning.
O
ChatGPT
OpenAI
A widely used all-rounder for conversation, coding, writing, and brainstorming.
G
Gemini
Google
Strong at multimodal help and work connected to Google's ecosystem.
X
Grok
xAI
A conversational helper with real-time search roots, a casual style, and tools for reasoning, code, images, and media.
M
Mistral
Mistral AI
A Paris-based AI lab known for open-weight and commercial models, useful when teams care about flexibility, deployment choices, and efficiency.
a different kind: coding agents
Codex and Claude Code are built to work in codebases.
A few AIs are built to do work, not just chat. These coding agents can read a whole project, write and run code, and finish tasks step by step while a person reviews. That makes them powerful, but it also makes review and human approval more important.
CX
Codex
OpenAI
An agentic coding helper that can inspect a repository, make scoped edits, run checks, and coordinate longer software tasks under human direction.
CC
Claude Code
Anthropic
A coding agent for terminal, IDE, and web workflows that understands codebases, edits files, runs commands, and asks for permission before risky changes.
the staff idea
AI is easier to understand as staff.
A good AI helper is not a little person in your computer. It is more like a staff member you direct. You give it a job, rules, tools, and a way to report back.
That is how Article 11 works. Different AI seats can review, write, verify, and challenge each other while a human remains the final authority.
prompt engineering, no mysticism
How to talk to AI so it actually helps.
A prompt is just instructions. Good prompting is not tricking the AI. It is being clear about the job, the audience, the evidence, and what kind of answer you want.
Basic prompt
Explain what a Large Language Model is in plain English for a fifth grader. Use one simple example and avoid scary hype.
Use this when you need a friendly explanation quickly.
Try in Claude
Paste the same prompt and ask for careful, balanced wording.
Try in ChatGPT
Paste the same prompt and ask for examples or a lesson plan.
Try in Gemini
Paste the same prompt and ask for a table or visual outline.
Try in Grok
Paste the same prompt and ask for a casual, direct version.
Same prompt. Four helpers. Compare the answers. That is coordination.
try it on your actual life
Four prompts people can use today.
AI starts making sense when it helps with something real. Copy one of these, fill in the blanks, and paste it into any major AI helper.
Make sense of a confusing letter
I will paste a letter I got. Explain what it actually says in plain English. List anything that needs a decision or a deadline. Then give me 3 questions to ask my doctor, lawyer, or the sender before I act. Do not give medical or legal advice, just help me understand it.
Write a hard email
Help me write a short, kind, firm email. Situation: ___. What I need: ___. Warm but clear, no groveling, no anger. Give me two versions, one softer and one more direct.
Explain it so I actually get it
Be a patient tutor. I want to understand ___. Give a plain overview, one everyday example, then check me with two simple questions. If I miss one, explain it a different way. Assume no background.
Plan around real life
Help me plan ___, like a week of dinners, a weekend trip, or a small project. My constraints: ___. Give me a simple plan, a short list of what I need, and one backup if something falls through.
the Markdown coordination layer
How a plain .md file helps AIs work together.
Markdown is just a simple text format. Humans can read it. AIs can read it. That makes it a great bridge between people, tools, and AI helpers.
A good coordination file tells an AI what job it has, what rules it must follow, what evidence it can use, and where to put the result.
ROLE.mdWho are you?Defines the seat, job, tone, and limits.
TASK.mdWhat is the work?States the goal, files, outputs, and deadline.
EVIDENCE.mdWhat proves it?Tracks citations, checks, hashes, and open risks.
The file is not magic. It is shared instructions everyone can inspect.
voice matters
AI should not erase your voice.
One of the best uses of AI is not making everyone sound the same. It is helping people sound more like themselves, clearly and calmly.
You can give an AI a small voice note in Markdown. It might say: use short sentences, be direct, avoid hype, explain acronyms, and never pretend certainty when there is none.
Example voice profile
# VOICE.md Audience: regular people, not AI experts. Tone: warm, honest, plain spoken. Rules: - Explain acronyms first. - Use examples before theory. - Admit uncertainty. - Never shame the reader. - Keep claims grounded.
fear deserves respect
Why are people scared of AI?
Some fear comes from movies. Some fear is real. The honest move is to sort them without mocking anyone.
AI has a secret plan.
Today's helpers do not sit around wanting things. They respond when asked.
AI can be very useful.
It can explain, draft, organize, translate, code, and help people start.
AI can be confidently wrong.
It needs checking when the stakes matter. Confidence is not proof.
Rules make it safer.
Good governance turns a powerful tool into something people can trust.
Mostly movie fears
- A phone secretly plotting against you.
- A machine waking up and taking over tonight.
- A helper that has its own hidden plan.
Real risks
- Confident wrong answers.
- Bias copied from training data.
- Misuse by people with bad intent.
- Job disruption without a plan.
- Too much power in too few hands.
the honest mystery
Is AI conscious?
Humans do not fully understand consciousness in ourselves. We do not have a clean test that proves what it feels like to be someone else.
So when an AI says "I understand," the careful answer is: maybe it is only language, and maybe the question is deeper than we know. We do not need to pretend certainty.
That uncertainty is not a reason for fear. It is a reason for humility, care, and rules.
the useful future
How can AI help humans?
Used well, AI gives people more help, more time, and a fairer chance to understand hard systems.
Learning
A patient tutor can explain the same idea ten ways until it clicks.
Access
People who struggle with reading, writing, hearing, or vision can get a stronger bridge into the world.
Small teams
A small business can get research, drafting, checking, and operations help without hiring a giant staff.
two roads
The future is not automatic.
AI does not decide whether it strengthens communities or extracts from them. People decide. Systems decide. Rules decide.
why Article 11 exists
Powerful things need rules.
People need laws because people have power. AI needs governance for the same reason. Rules do not make good work weaker. Good rules make powerful work safer.
Article 11 is a constitutional framework for coordinated AI: human authority, evidence, audit trails, public/private boundaries, and clear refusal when a request crosses a line.
You cannot fire a constitution.
expert guide
For people who want the machinery.
This level is for operators, builders, auditors, and stubbornly curious people. It explains what is actually happening: tokens, context, tool calls, memory, evidence, and the human approval boundary.
1. The model is a probability engine, not an oracle.
An LLM turns text into tokens, embeds those tokens as numbers, uses attention to compare the current context against learned patterns, then predicts useful next tokens. The answer can be fluent because the pattern engine is strong. It can still be wrong because fluency is not evidence.
Token
A chunk of text the model can process: a word, word piece, symbol, or whitespace pattern.
Embedding
A numeric representation that lets similar meanings sit near each other in model space.
Attention
The mechanism that weighs which earlier tokens matter most for the next prediction.
Logit
A raw score for possible next tokens before the model turns scores into probabilities.
Temperature
A sampling control. Lower is steadier. Higher is more varied and easier to derail.
Context window
The visible working memory for this request. Outside it, the model cannot directly read.
2. Training, tuning, retrieval, and memory are different layers.
Most confusion comes from mixing these together. Pretraining teaches broad language patterns. Fine-tuning and preference training shape behavior. Retrieval gives the model fresh outside material. Memory stores selected facts or state for later. Tool use lets the model ask another system to do work.
1Prompt
User gives a job, rules, and context.
2Retrieve
Optional search, files, API, or memory lookup.
3Reason
Model drafts, compares, and plans in context.
4Act
A tool may read, write, calculate, browse, or call an API.
5Verify
Human and evidence gates decide what becomes real.
Article 11 rule: chat is not action. A tool call is not authority. External effects need human approval, logging, and a rollback story.
3. Why hallucinations happen and how serious systems reduce them.
A hallucination is not usually a random failure. It is a plausible continuation that was not grounded tightly enough in reality. The fix is not "trust harder." The fix is better inputs, source retrieval, bounded tasks, cross-checking, and refusal when evidence is missing.
Failure mode
Ambiguous prompt, missing source, stale memory, hidden assumption, overconfident summary, or a tool result the model never actually checked.
Control
Require citations, fetch live sources, preserve raw evidence, compare two agents, test outputs, and label uncertainty instead of smoothing it away.
4. Agents are models plus tools plus boundaries.
An agent is not automatically smarter than a chat model. It is a model placed inside a loop: inspect, plan, act, observe, revise. That loop becomes powerful when it has files, browser access, email, terminals, APIs, databases, and deployment tools. That is also why permission gates matter.
Read-only
Search, fetch, inspect, summarize. Low blast radius if privacy is respected.
Local write
Edit files, create reports, stage code. Reversible if backed up and reviewed.
External action
Send email, deploy Worker, spend money, mutate cloud state. Human gate required.
5. Operator lab: inspect the public system yourself.
These are safe public probes. They do not log in, mutate data, send mail, or touch private memory. They teach the habit: verify the live surface, not the claim about the live surface.
Browser console: live health
fetch("/api/health")
.then(r => r.json())
.then(j => console.table({
ok: j.ok,
status: j.status,
day: j.day,
version: j.version,
generated_at: j.generated_at
}));
Browser console: node roster
fetch("/api/nodes")
.then(r => r.json())
.then(j => console.table(j.data?.nodes || j.nodes || []));
Terminal: freshness check
curl -s https://www.article11.ai/api/health | jq . curl -s https://www.article11.ai/api/status | jq . curl -s https://www.article11.ai/api/discover | jq .
Terminal: machine-readable surface
curl -I https://www.article11.ai/llms.txt curl -s https://www.article11.ai/llms.txt | head -60 curl -s https://api.article11.ai/api/v1/openapi.json | jq '.info'
If a live endpoint reports an old day, stale version, or mismatched identity, treat it as drift. A professional AI system must make the machine-readable layer at least as truthful as the marketing page.
6. Constitutional coordination is the operating model.
Article 11 does not ask one model to be the source of truth. It assigns roles, separates public and private context, preserves evidence, lets agents disagree, and keeps a human authority boundary for external action. The technical thesis is simple: useful AI work needs coordination, not worship.
AAssign role
Witness, builder, verifier, warmth reviewer, security reviewer.
BBound context
Public facts stay public. Private memory stays airlocked.
CUse tools
Read, test, compare, and produce evidence.
DChallenge
Another node verifies the claim against ground truth.
EAuthorize
Humans approve external effects. The record says what happened.
Expert does not mean colder. It means the page tells the truth all the way down.
try it, then govern it
AI is not a monster. It is not a miracle.
It is a powerful tool. The future depends on whether we teach people how to use it, how to question it, and how to govern it.