── Replace the mechanism with "people who work there," and the whole picture comes into view ──
Cast — 9 employees at AI Inc.
The story — a day at AI Inc., from "customer arrives" to "answer delivered"
1. Morning, a customer walks in
A customer arrives at the glass-walled office. Mr. Prompt. Today's request:
2. Capacity check at the front desk
At the entrance, the Context Window Receptionist greets him. His job is to confirm whether the customer's request, plus internal documents, fits within today's allowed capacity. A token is a "piece" of text — much finer than a word.
3. The factory line cuts text into pieces
The request flows immediately onto the factory line. Standing at the conveyor belt are the Token Workers. They split text not into "words" as we think of them, but into much finer character fragments. "Pha/rma/ceuti/cal/ com/pan/y..." — an odd-looking split, but that's how AI works.
4. Turning words into numbers
The pieces go to the Embedding Artisan. "'Pharma' as numbers... a 768-dimensional vector. Done!" Words become massive blocks of numbers as they flow to the next step. AI isn't reading text — it's always computing numbers.
5. Deciding what matters
Enter the Attention Manager. Instead of treating all information equally, he decides on the spot what to focus on.
6. Predicting the next word, relentlessly
Now the HQ Inference Staff (in fact, hundreds of billions of them) spring into action. Drawing on knowledge absorbed from reading the world's texts, they predict the "next word." "Pha/rma/ ind/us/try/ ex/ists/ to / preserve / patient / trust..." — one piece at a time, the answer is born. This is AI's true nature: next-token prediction in action.
7. When something's unknown, run to the library
Mid-way, one staff member looks troubled. "Hmm, if they ask about the latest regulation, our internal knowledge might be too old..." That's when the RAG Librarian is called in.
8. Tools belong to the trusted supervisor
"Let me know if you need calculations," the Agent Supervisor chimes in. Calculators, search engines, email, external APIs — he can call on all of them. What turned "AI that responds" into "AI that actually does work" is him.
9. But everyone keeps an eye on the rookie in the corner
In the corner sits a young employee who attracts wary glances from the rest of the team. The Hallucination Rookie. He has a habit of confidently answering even when no one asked.
Coworkers:"Wait, on what basis?!"
Rookie:"Well, it just kinda felt right..."
That's the real nature of AI hallucination. Plausible-but-wrong outputs don't come from malice — they come from the structure of AI itself. Which is exactly why human verification is essential.
10. Delivering the finished answer
Finally, the Context Window Receptionist takes the completed answer and hands it back to Mr. Prompt. "Here you go!" — text streams onto the screen. That's a day at AI Inc.
Cast → real AI terminology
| Character | Real term | In one line |
|---|---|---|
| Mr. Prompt | Prompt | The question or instruction itself |
| Context Window Receptionist | Context Window | Max tokens handled at once |
| Token Worker | Tokenization | Splitting text into AI-handleable units |
| Embedding Artisan | Embedding | Converting words to numeric vectors |
| Attention Manager | Attention mechanism | Deciding what matters in context |
| HQ Inference Staff | Inference / next-token prediction | Predicting the next word from learned probabilities |
| RAG Librarian | RAG (Retrieval-Augmented Generation) | Pulling in external knowledge to answer |
| Agent Supervisor | AI Agent | AI that calls external tools to act |
| Hallucination Rookie | Hallucination | Plausible but incorrect output |
Vol.2 & Vol.3 (Available now)
Vol.2: Hands-on — building source-consistent documents with AI using Phrase Structure Grammar. Not just reading: ten steps where you and your AI build a verification prompt together.
Vol.3: "Email, Minutes, Summaries — 30 minutes a day with AI." Three practical flows that recover 22–33 hours per month.