0 Level 0 · Zero Knowledge

What is an AI,
really?

Five things to un-confuse before anything else makes sense. No jargon walls, no hype, just the actual mental model of what's happening when you hit send.

6 min read Beginner No prerequisites
01 — Definitions

What is an AI, really?

You type something into ChatGPT. Maybe a recipe question, maybe an email draft, maybe something you were too lazy to Google. In milliseconds, out comes an answer. Fluent, coherent, sometimes better than what you'd write yourself.

But what actually happened in those milliseconds? Not "AI thinking." Not "AI understanding." Not magic. What happened is the model loaded everything it ever learned from the internet, roughly 15 trillion tokens of text (that's about all of Wikipedia written 50,000 times over, from Reddit threads to romance novels to programming forums), and then it guessed. Literally guessed the next most likely word, given the context so far.

The guess is often right, which is why it looks smart. Sometimes the guess is wrong, and it doesn't know it's wrong, which is why it can look stupid with total confidence. Level 0 is for understanding why both of those things happen at the same time. First, let's define three words people use interchangeably but mean very different things.

AI (Artificial Intelligence) is the broadest term. It covers the entire field, any system that does something that usually requires human intelligence. Your phone's face unlock is AI. Spotify's recommendation engine is AI. A chess bot is AI. The term is so broad it's almost useless without context.
ML (Machine Learning) is a subset of AI. Instead of programming explicit rules ("if this, then that"), you feed data to an algorithm and let it figure out the patterns itself. Traditional programming: you write the rules. Machine learning: the machine writes the rules from examples.
LLM (Large Language Model) is a specific type of ML. It's trained on massive amounts of text and learns to predict what word comes next. That's it. ChatGPT, Claude, and Gemini are all LLMs. They don't "understand" anything. They're extremely good at predicting plausible text.
AI — the whole field ML — learn from data, not rules LLM — a text-prediction machine
When someone says "AI" in 2026, they usually mean LLMs. But knowing the difference helps you understand what you're actually dealing with.
⚡ Try this now
Open ChatGPT or Claude and type:
Explain the difference between AI, ML, and LLM in one sentence each, using a kitchen analogy.
Read the result. Does the analogy hold up? If you can repeat it back to a friend, you now actually know the difference.
02 — Training

How does AI "learn"?

An LLM doesn't learn the way you do. You read a book, understand the concepts, and can apply them in new situations. An LLM reads the entire internet and memorizes statistical patterns.

Here's the simplified process:

1

Training data

The model is fed billions of text documents: books, websites, forums, code, everything. It doesn't "read" them. It processes them as sequences of tokens (more on that later).

2

Pattern recognition

The model looks at every sequence and asks: "Given these words, what word is most likely to come next?" After seeing billions of examples, it builds an incredibly complex statistical model of language.

3

Fine-tuning

After the initial training, the model is further trained on curated examples — often with human feedback — to make it more helpful, less harmful, and better at following instructions.

The result: a machine that's really, really good at predicting what text should come next in any given context. It doesn't "know" things. It doesn't "think." It generates text that statistically matches the patterns it learned.

This is why AI is sometimes brilliant and sometimes confidently wrong. It is optimized for plausibility, not accuracy. Those sound similar until the model gives you an answer that reads perfectly and is completely false.
03 — Hallucination

Why does AI sometimes lie?

It doesn't lie. Lying requires knowing the truth and choosing to say something else. AI doesn't know truth from fiction. It predicts plausible text.

This is called hallucination: the AI generates something that sounds right but isn't. Tap a prompt below to watch it happen:

live demo · watch the model predict

Why this happens:

  • The model was trained to produce plausible text, not factual text
  • It has no mechanism to verify truth, only to predict patterns
  • When it's "uncertain," it doesn't say "I don't know." It generates its best guess
  • The more specific the question, the higher the chance of hallucination

How to deal with it:

  • Never trust AI output for critical facts without verification
  • Ask for sources, then check those sources yourself
  • Use AI for structure, reasoning, and drafting, not as a factual authority
  • If something sounds too specific and too confident, double-check it
Hallucination is a fundamental characteristic of how these models work, not a bug that will be "fixed." Understanding that single fact changes how you use AI, from "ask and trust" to "ask and verify."
⚡ Try this now
Type into ChatGPT:
Who first invented the telephone? Give me a source.
Look closely. Is there a source? Does the source actually exist? Google it. Models sometimes invent citations that look perfectly real but point to nothing.
04 — The models

ChatGPT vs Claude vs Gemini

They're all LLMs trained on similar data with similar techniques. But they're built by different companies with different priorities, and those differences matter. Tap to compare:

ChatGPT

OpenAI
  • The most well-known. Massive user base.
  • Strong at general tasks, creative writing, coding.
  • Tends to be more "chatty" and agreeable.
  • Ecosystem: plugins, GPTs, DALL-E, web browsing.
  • Weakness: can be overly verbose, sometimes sycophantic.

Claude

Anthropic
  • Built with a focus on safety and helpfulness.
  • Strong at nuanced reasoning, long documents, careful analysis.
  • Better at saying "I don't know" when uncertain.
  • Longer context window (can process more text at once).
  • Weakness: more cautious, sometimes refuses reasonable requests.

Gemini

Google
  • Deeply integrated with the Google ecosystem.
  • Strong at multimodal tasks (text + images + video).
  • Good at research and factual queries (backed by Google Search).
  • Weakness: inconsistent quality, sometimes feels less polished.

How to pick:

  • Creative tasks → ChatGPT or Claude
  • Long documents → Claude
  • Research with sources → Gemini
  • Coding → all three are strong, preferences vary
  • Daily driving → try all three, pick what feels right
The honest truth: in 2026, the differences are smaller than people think. What matters more is how you use them (prompt engineering) and what tools you connect them to (harness engineering). A great prompt on a "weaker" model often beats a bad prompt on the "best" model.
05 — Economics

Tokens — the currency of AI

Every time you interact with an LLM, you're spending tokens. Understanding tokens is understanding the economics and limits of AI.

What is a token?

A token is a chunk of text, roughly 4 characters in English, or about 3/4 of a word. The sentence "I love programming" might be tokenized as: ["I", " love", " programming"], which is three tokens.

But tokenization isn't always intuitive:

  • "Don't" might be one token, or two: ["Don", "'t"]
  • Rare words get split: "tokenizer" → ["token", "izer"]
  • Code, different languages, and special characters tokenize differently

Why tokens matter:

Tokens used 500K
Price per 1M tokens $30
$15.00
500,000 tokens @ $30 / 1M
  • Context window. Every model has a maximum number of tokens it can process at once, its "context window." GPT-4o: 128K tokens. Claude Sonnet 4: 200K tokens. Gemini: up to 2M tokens. Once you hit the limit, older messages get cut off. This is why AI "forgets" in long conversations.
  • Speed. More tokens = slower response. A 10-token answer comes back faster than a 1,000-token essay.

Practical implications:

  • Shorter prompts = cheaper and faster
  • Asking for concise output saves money
  • Long conversations eat tokens because the entire history is re-processed each turn
  • When a model seems to "forget," it's usually because older messages fell outside the context window
Think of tokens like phone minutes in the early 2000s: you have a limited budget, and every interaction costs something. Being mindful of token usage is one of the most practical skills in working with AI.
⚡ Try this now
Type anything into ChatGPT, then ask:
How many tokens did you use in this response?
Now you know why AI sometimes "forgets" earlier parts of a long conversation. The context window literally ran out of room.
06 — Watch out

What can go wrong at this level

You're just starting out. These are the four mistakes people make most often at this stage, and how to sidestep each one.

1. Trusting a confident answer blindly

You ask something, the AI answers with total confidence, and you believe it. The answer might still be wrong. This isn't the AI lying. It has no concept of "right" or "wrong." It's only guessing what sounds right.

How to avoid: If the AI gives you a fact (a number, a name, a date, a link), verify it. Google it. Check another source. Never trust a specific claim without checking it.
2. Confusing AI, ML, and LLM

You use these terms interchangeably. "AI will take our jobs." Which AI? ML is not LLM. LLM is not AGI. If you can't tell the difference, you can't discuss this topic intelligently.

How to avoid: Re-read section 01. AI is the broad umbrella. ML learns from data. LLM predicts the next word. Three genuinely different things.
3. Thinking the AI "understands" because the output is good

This is the most dangerous one. You read AI output that's coherent and well-structured, and you conclude "it understands." It doesn't. It's pattern matching, like autocomplete that happens to be very good.

How to avoid: Ask yourself: "If a human wrote this, would I trust it?" If yes because they're genuinely knowledgeable, that's valid. If yes because it sounds knowledgeable, that's a red flag.
4. Copy-pasting without re-reading

You ask the AI to write something, then paste it straight into an email or document. Without re-reading, without checking, without editing.

How to avoid: Always re-read AI output before you use it. Edit it. Adapt it. The AI produces the draft. You finalize it.
06 — What you now know

What you should know after Level 0

If you've read this far, you now understand the foundations. Tap each item as it clicks, and track your own literacy:

You're not an expert. But you're no longer guessing. You have a mental model of what's happening when you type a prompt and get a response. That mental model is the foundation for everything that comes next.