AI Engineering Studio

Appearance

How LLMs Actually Work (no math) ​

📝 Context ​

Enough of a working model to be dangerous in a meeting — without a single equation. The goal isn't to make you an ML researcher; it's so that when a customer asks "why did it make that up?" you can answer truthfully and calmly instead of changing the subject.

A large language model does exactly one thing: given some text, it predicts the next chunk of text. It has read an enormous amount of writing and learned the statistical shape of language — which words tend to follow which. Given a prompt, it predicts the most likely continuation, one token at a time, feeding each prediction back in to predict the next.

flowchart LR
  P["Your prompt: the capital of France is"] --> M["Model predicts next token"]
  M --> T["Paris"]
  T --> M2["feed it back, predict again"]
  M2 --> D["end of sentence: done"]

A "token" is just a piece of a word. Models read tokens — roughly 3–4 characters, about ¾ of a word on average (illustrative; varies by language). It matters because you're billed per token and the context window is measured in tokens, not words. "$0.30 per million tokens" is roughly a million ¾-words.

Everything impressive an LLM does — summarizing, coding, reasoning — is this same next-token prediction, scaled up. It isn't looking anything up or consulting a database. It's continuing the pattern. You don't need the transformer math to be effective; you need the behavior.

🎯 Four Mental Models That Explain Almost Every Behavior ​

1. The context window is its only working memory ​

The model can only "see" what's in the current prompt — the context window. Everything outside it doesn't exist to the model. Previous conversation, your documents, yesterday's correct answer: unless it's in the window right now, it's gone. This is why RAG exists — it puts the relevant documents into the window at the moment you ask.

"The model has no memory between calls — every request starts from a blank slate. So 'it forgot what I told it' isn't a bug, it's the architecture. We design around it by deciding what to put in the window each time."

2. The training cutoff means it doesn't know recent or private things ​

The model learned from data up to a certain date — its training cutoff — and from public text. It has never seen your internal wiki, last week's news, or anything proprietary. Ask it about those and you get either an honest "I don't know" or a confident fabrication. Which brings us to the question every customer asks.

3. Hallucination is the system working as designed, not a malfunction ​

Because the model's job is to produce plausible continuation, it produces plausible-sounding text even when it has no grounding for it. It doesn't know what it doesn't know — there's no internal "confidence meter" before it invents a citation. A hallucination is the model doing what it always does (predict likely text) in a case where likely text happens to be false.

Say it like this

"It's not lying and it's not broken — it's a prediction engine, and sometimes the most fluent-sounding answer isn't the true one. The fix isn't a smarter model, it's grounding: we feed it your actual documents and tell it to answer only from those and say 'I don't know' otherwise. That's what turns a demo into something you can trust."

4. Temperature is a creativity dial, not an accuracy dial ​

When the model picks the next token, temperature controls how much it favors the single most-likely option versus sampling less-likely ones. Low → repetitive, deterministic. High → varied, creative, more prone to drift. It's a knob for style, not correctness — turning it down reduces randomness but does not make the model know more.

🧩 Worked Scenario: A Demo That Hallucinated On Stage ​

You're demoing a support assistant. Someone asks about a product feature that shipped after the model's training cutoff and isn't in the retrieved docs. Asked a direct question, the model produces a confident, detailed, wrong answer. The room goes quiet.

  • What happened — no grounding in context for that feature, so the model predicted plausible text, and plausible ≠ true.
  • What it wasn't — not a "dumb model." A frontier model does the same thing; size doesn't fix ungrounded questions.
  • The real fix — retrieval that covers the feature, plus a system instruction to refuse when context is missing.
  • The recovery line — name it honestly in the room; it builds trust faster than pretending it didn't happen.

⚠️ Gotchas ​

  • "A bigger/smarter model will fix the hallucinations" — no; an ungrounded question gets a confident guess from any model. Grounding fixes it.
  • "Turn the temperature down for accuracy" — temperature controls randomness, not knowledge.
  • "It remembers our last conversation" — only if you put that history back in the context window.
  • Explaining transformers/attention to a customer — wrong altitude; talk behavior, not architecture.