World 1-2: underground

Décio Dalke Jr. — Arquiteto de negócios. Sócio-gerente no ecossistema MitUP (mitup.pt).

World 1 of AI series — level 2 of 5: World 1-1 · World 1-2 · World 1-3 · World 1-4 · World 1-5

The second level of Super Mario’s first world takes place underground. The screen goes dark, the sky disappears, and what used to be scenery becomes the ceiling. You already know how to jump, you already know what the mushroom does. Now you need to understand the terrain.

With AI, World 1-2 is the same thing. World 1-1 covered what it is (an assistant), what you want to use it for and how you’ll check what you get. This level goes down one floor: what happens underneath, and why that decides whether the assistant will be junior or senior in your hands.

AI only knows what’s in front of it at the moment it answers: the context window. That window is measured in tokens, it has a size and it has a cost, and more context isn’t better context. When data is missing, the model tends to fill the gap with something plausible. That’s hallucination.

Why do data, context and structure matter so much?

A quick explanation is in order — deliberately simplified — of how these models work.

A language model was trained on an enormous amount of text to do one thing: predict the next piece of text. It does this so well that it seems to reason, and in many cases the result is indistinguishable from reasoning. But it doesn’t know your company. It doesn’t know who your customer is, what was decided yesterday, which of the spreadsheets is the right version. Everything it knows about your case is what’s in front of it at the moment it answers.

That “in front of it” has a name: the context window. Anthropic, which develops Claude, describes it as the model’s working memory — something different from everything it was trained on [1]. Picture a brilliant consultant who arrives every morning remembering nothing from the day before. All they know is what’s in the folder you leave on their desk. If the folder is empty, they work with what they know about the world in general. And they fill in the rest with whatever seems most likely.

That’s why these three elements carry so much weight:

• Data is the facts: numbers, documents, history. Without it, the model works with the world’s average, not with your reality.
• Context is the situation: the goal, who it’s for, what has already been decided, what must not happen. Without it, the answer can be right and useless.
• Structure is where each thing lives and who can touch what. What’s the source of truth? If there are three versions of the same document, which one counts?

Structure is the one that shows the least and costs the most. When several agents are working at the same time and sharing information, the question stops being “what’s the best prompt” and becomes “who can do what, and where is the truth”. Those are systems architecture requirements. It’s also why AI doesn’t solve problems by magic.

What are tokens, and why isn’t more context better context?

The model doesn’t read words. It reads tokens — pieces of text, sometimes a whole word, sometimes a syllable or a punctuation mark. Everything is measured in tokens: what you write, what it answers, what it reads from documents. Tokens are what you pay for, and tokens are what fill up the window.

A detail few people notice: in a conversation, the model doesn’t “remember” the previous message. With each new message, the entire conversation is read again — every question and every answer keeps piling up in the window [1]. A long conversation gets more expensive with every turn.

And more context isn’t better context. Anthropic’s own documentation warns that as the volume grows, accuracy drops [1]. A Chroma study tested 18 models in 2025 and concluded that they don’t use context uniformly: performance becomes less reliable as the input grows, even on simple tasks [2]. Stanford researchers had already shown that, in a long context, information in the middle is what gets lost most; the beginning and the end are used better [3].

In practice, dumping everything into the conversation “just in case” is handing the consultant with no memory the entire archive instead of the right folder. They’ll read it. And they’ll get lost in the middle.

There’s one particularly dangerous moment. When the conversation fills up, the tools summarise what came before to make room. The summary isn’t the conversation. In my operation, the worst errors in a work session all happened after one of these summaries: the assistant kept treating what it “remembered” as if it were what was written in the files. Two rules came out of that. Decisions get recorded in writing on the spot, not “at the end of the session”. And after an automatic summary, the source is re-read before saying anything about what has already been decided.

Why does AI hallucinate?

Hallucination is when the model states something false with complete confidence. A figure that doesn’t exist, a made-up reference, a plausible and wrong number.

It isn’t a mysterious defect. In 2025, OpenAI researchers published a very straightforward explanation [4]. Part of it comes from training itself: predicting the next word works very well for patterns (grammar, style) and poorly for rare facts, like a specific person’s birthday. The other part comes from how models are evaluated: if the score only counts correct answers, guessing pays. Guess someone’s birthday and you have a one-in-365 chance of getting it right; say “I don’t know” and you’re guaranteed zero. The model learns what the report card rewards.

Translated to the business: your assistant was, in a way, brought up never to leave a question unanswered. If you don’t give it the data, it fills the gap. And it fills it well — with something that looks right.

The two defences aren’t technological. The first is to give it the data (back to data, context and structure). The second is to allow the assistant not to know: if information is missing, stop and ask. And, in any case, check — which was the question in World 1-1.

What I’ve learned running this every day

I’ve been using AI in running my company since March 2025, with agents doing real work. Every rule in my method was born from something that went wrong. The main ones:

• Never make things up. Nothing goes into a deliverable without a source. If data is missing, stop and ask. A document that’s 20% filled in is delivered 20% filled in, saying what’s missing and who it’s being requested from.
• Prove, don’t declare. “Done”, “right”, “identical” only with the proof alongside: the count, the comparison, the open file. Confidence isn’t proof.
• Extreme numbers get checked before they’re reported. When the result is almost everything or almost nothing, the first suspect is the method, not the world. One of our audits said 73 out of 112 people had no account in a system. It seemed like too many. It was false: the comparison used full names, and the system appended a suffix to the name. The real number was 42.
• One piece of information, one home. Each fact lives in one place only. Other documents point to it; they don’t copy it. Every copy is a chance to diverge.
• One writer at a time. Two agents editing the same document at the same time wipe out each other’s work. It has happened here.
• Disagree, don’t agree out of habit. An assistant that only says yes is worse than none: it gives false security.

None of these rules is about the tool. All of them are about method.

Underground, Mario doesn’t need a new jump. He needs to see where he’s stepping.

Quick questions

What is the context window? The model’s working memory: everything it has in front of it at the moment it answers. Whatever isn’t there, it doesn’t know about your case.

What are tokens? The pieces of text the model reads and writes with. Tokens are what you pay for, and tokens are what fill up the window.

Why does AI hallucinate? Because it was trained to predict plausible text and evaluated in a way that rewards guessing over saying “I don’t know”. The defences: give it the data, allow the assistant not to know, and check.

References

1. Anthropic — Context windows (Claude documentation). https://platform.claude.com/docs/en/build-with-claude/context-windows — accessed 2026-09-18.
2. Chroma — Context Rot: How Increasing Input Tokens Impacts LLM Performance, 14 July 2025. https://www.trychroma.com/research/context-rot — accessed 2026-09-18.
3. Liu, N. F. et al. — Lost in the Middle: How Language Models Use Long Contexts. Transactions of the ACL, 2024 (arXiv, 2023). https://arxiv.org/abs/2307.03172 — accessed 2026-09-18.
4. OpenAI — Why language models hallucinate, 5 September 2025. https://openai.com/index/why-language-models-hallucinate/ (full paper: https://arxiv.org/abs/2509.04664) — accessed 2026-09-18.