Try Shots. Free.

Everything you see AI do today (write an email, generate an image, answer a question, write code) is the result of seventy years of mathematics built by millions of scientists across the world. It is not a genie. It is not magic. It is a very sophisticated pattern-matching engine that predicts the next most likely word, pixel, or token. That's it. And once you understand that, you stop being afraid of it.

When AI gives you a wrong answer confidently, it's not lying to you. It's doing math that went slightly off. When you understand that, you can fix it instead of feeling fooled by it.

Open any AI tool you use: ChatGPT, Claude, Gemini, it doesn't matter. Ask it: "Explain to me how you work, like I'm a curious ten-year-old." Then ask a follow-up: "Where are you most likely to get things wrong, and why?" Share one thing that surprised you.

What's one thing you believed about AI before today that feels different now?

The Illustrated Transformer, Jay Alammar

You now know AI is pattern-matching, not magic. So here is your challenge: find one moment this week where you caught AI being wrong: a confident mistake, an odd phrasing, an answer that felt off. Screenshot it. Then ask AI why it got it wrong. Share what it said. This is not a gotcha exercise. It is calibration. The people who use AI best are the ones who know its failure modes as well as its strengths.

AI didn't appear overnight. It started in the 1950s with simple models called perceptrons that could barely tell a circle from a square. Then came the AI winters, decades where funding dried up because the technology just wasn't working. Then breakthroughs: neural networks in the 80s, deep learning in the 2010s, transformers in 2017, and ChatGPT in 2022, which put everything into the hands of ordinary people. You are living at the peak of seventy years of stubbornness and curiosity.

When you know the history, you understand why AI feels like it arrived suddenly. It didn't. It was being built quietly for decades. And knowing that makes you trust it more intelligently, not blindly.

Ask any AI tool: "Give me a timeline of the five most important moments in AI history, and explain why each one mattered." Then ask: "What do you think comes next in the next five years?" Screenshot the answer and share it with your reaction.

Which moment in AI history surprised you the most, and why?

AI took 70 years to reach your phone. Think about one skill or craft in your own life (cooking, writing, managing people, designing) that took years to compound quietly before it clicked. Open an AI tool and have a conversation about that journey. Ask it: "How does the way humans develop expertise compare to how AI models are trained?" Then write two sentences: what's similar, and what's different. You'll understand both better by the end.

When you type a sentence to an AI, it doesn't see words the way you do. It breaks your sentence into small chunks called tokens, sometimes a whole word, sometimes just part of one. It then converts each token into a number, and that number lives inside a vast mathematical space called an embedding. Words that mean similar things live close together in that space. "Cat" and "kitten" are neighbours. "Cat" and "invoice" are far apart. This is how AI understands meaning: not through dictionaries, but through mathematical proximity.

This is why punctuation matters, why rephrasing a prompt can get you a completely different answer, and why longer, more specific prompts often work better. You're not typing to a person. You're navigating a mathematical space with words.

Ask any AI tool: "What is tokenisation and show me how you would break this sentence into tokens: 'She vibes with AI every single day.'" Then ask: "Why does rephrasing a prompt change your answer?" Share what you learned.

Does knowing AI reads numbers instead of words change how you'll write prompts going forward?

Tokenisation explained, OpenAI Tokeniser tool

You know AI sees tokens, not words. Now test it in your own work context. Take a paragraph you wrote recently (an email, a message, a document) and paste it into the OpenAI tokeniser linked above. Count the tokens. Then rewrite the same paragraph more concisely and count again. Now ask Claude or ChatGPT: "Which version of this paragraph will give you better context to respond to?" You're starting to think like a prompt engineer.

Here is the single most important thing to understand about how AI generates text: it is predicting the next most likely token, one at a time. Every single word you see in an AI response is the result of a calculation: given everything said so far, what word comes next? There is a mathematical function called softmax that turns a huge list of possible next words into probabilities. The AI picks from those probabilities. That's it. No understanding. No feelings. Pure prediction.

This is exactly why AI hallucinates. It isn't lying. It is confidently predicting what words should come next, even when it has run out of real knowledge to draw from. Once you understand this, you stop being shocked when it makes things up, and you start knowing when to verify.

Ask any AI: "Explain next-token prediction to me like I've never heard of it. Then give me an example where this prediction process goes wrong, a hallucination, and explain why it happened mathematically." Share the hallucination example.

Now that you know AI is predicting word by word, does that change how much you trust it? What will you verify from now on?

How Large Language Models Work, Stephen Wolfram

AI hallucinates because it predicts. You have intuition that also sometimes predicts wrong. Think of one domain in your work or life where you make quick calls based on pattern recognition: reading a room, estimating a timeline, judging a person's mood from a message. Ask AI: "Help me build a checklist to verify my instinct in [that domain] before I act on it." Compare how you handle your own uncertainty versus how AI should handle its uncertainty. The insight here is about knowing when to pause and verify, in yourself and in the tool.

Before AI can predict anything, it has to learn. This happens in stages. First, pretraining: the model reads a vast portion of the internet and learns the patterns of human language. This takes months and costs millions. Then fine-tuning: the model is shaped to be helpful and safe, using human feedback. Real people rated responses as good or bad, and those ratings taught the model how to behave. This process is called RLHF, Reinforcement Learning from Human Feedback. Every thumbs up and thumbs down you've ever given an AI has been part of training future models.

You are not just a user of AI. You have been, without knowing it, a contributor to how it thinks. And understanding this helps you see AI not as an oracle, but as a mirror of human knowledge, with all its biases and blind spots included.

Ask any AI: "Explain the difference between pretraining and fine-tuning. What kind of biases could creep in during each stage?" Then ask: "What is RLHF and how did human feedback shape how you respond?" Reflect on one bias you've noticed in AI responses and share it.

If AI learned from human feedback, what do you think it learned about how people communicate, ask questions, and seek help?

Reinforcement Learning from Human Feedback, Anthropic

AI was fine-tuned on human feedback. You've been fine-tuned by your experiences too. Think of one strong professional opinion you hold: how meetings should run, how feedback should be given, how projects get done. Ask AI the same question and compare its answer to yours. Then ask: "What life experiences or data might have shaped your answer differently from mine?" You're reverse-engineering bias in AI and in yourself simultaneously. That's the most human use of AI I can think of.