CrackKit
0% complete
AI Engineering: LLMs, RAG & Agents

How LLMs actually work (the honest mental model)

You don't need a PhD to build with LLMs — but you do need an accurate mental model. Most production bugs in AI apps come from developers holding the wrong one.

The honest one-liner

An LLM is a function that takes a sequence of tokens and outputs a probability distribution over the next token. That's it. Everything else — chat, reasoning, agents — is scaffolding built on repeatedly sampling "what token comes next?"

"The capital of France is" LLM " Paris" 92% " the" 3% " a" 1% generation = sample next token → append → repeat until stop

What "training" gave the model

  • Pretraining — predict the next token across trillions of tokens of text. Result: a compressed statistical model of language, facts, code, and reasoning patterns.
  • Post-training (instruction tuning + RLHF/RLAIF) — teach it to behave like an assistant: follow instructions, refuse harm, format answers. This is why "chat" models answer questions instead of continuing your sentence.

Consequences you must design around

Model propertyEngineering consequence
Predicts plausible text, not truthHallucinations — plausible ≠ correct. Ground with RAG (section 4) or tools
No memory between callsYou manage state: send conversation history every request
Fixed training cutoffFresh data must arrive via the prompt or tools
Sees tokens, not lettersStruggles with character-level tasks ("count the r's in strawberry")
Sampling is stochasticSame input can give different outputs — design for it, or pin temperature 0

The mental-model test: if a feature of your app assumes the model "knows" or "remembers" something you didn't put in the current request, the design is broken. The prompt is the entire universe the model sees.

Why they still feel smart

Scale. With enough parameters and data, next-token prediction forces the model to internalize grammar, world knowledge, code semantics, and reasoning-shaped patterns — because predicting text written by reasoning humans requires approximating reasoning. "It's just autocomplete" undersells it; "it thinks like a person" oversells it. Hold both.