← Learn LLMs path
Stage 1 of 6Beginner·2–3 hours

What an LLM Actually Is (and Isn’t)

Before prompting tricks or code, get the mental model right: an LLM predicts text, works inside a limited window, and can be confidently wrong. Everything else follows from those three facts.

The idea

A large language model (LLM) is a program trained on a very large amount of text to do one thing extremely well: given some text, predict what text should come next. Chat, summarising, translating, writing code, answering questions — all of it is that one skill applied cleverly. When you type a question, the model is not "looking up" an answer; it is generating the most plausible continuation, one small chunk (a token) at a time.

Three consequences matter for everyone. First, the model only knows what was in its training data plus whatever you put in front of it, so it can be out of date or simply not know your company’s facts. Second, it works within a context window — a limit on how much text it can consider at once — so very long documents or conversations get cut off or forgotten. Third, because it produces plausible text rather than verified facts, it can state something false with total confidence. This is called a hallucination, and it is not a bug that will be fully fixed; it is a property of how the tool works, so you verify anything that matters.

What LLMs are excellent at: transforming text (summarise, rewrite, translate, change tone), drafting (emails, plans, code), explaining (at any level you ask for), and extracting structure from mess (turn notes into a table). What they are weak at without help: precise arithmetic, current events, your private data, and anything requiring certainty. Modern products bolt on tools — web search, calculators, file access — to cover exactly those gaps, which is where later stages of this path go.

Vocabulary

Token
The chunk of text a model reads and writes in — roughly ¾ of a word in English. Pricing and limits are counted in tokens.
Context window
The maximum amount of text (your messages + its replies + any files) the model can consider at once.
Hallucination
A confident, fluent, wrong answer. Always possible; more likely on obscure facts, numbers and citations.
Training cutoff
The date after which the model has no knowledge unless it searches the web or you tell it.
Model
A specific trained system, e.g. Gemini 2.5 Pro, Claude Sonnet 4.5, GPT-5. Each provider offers several sizes.

No-code walkthrough · non-developers

See the strengths and the limits in 20 minutes

For: Anyone with a free ChatGPT, Claude or Gemini account

  1. 1Open ChatGPT, Claude or Gemini (all have a free tier). Paste a long email or article you have and ask: "Summarise this in 3 bullet points for someone who has 30 seconds."
  2. 2Now ask it to rewrite the same text three ways: "as a friendly text message", "as a formal notice", "for a 10-year-old". Notice it is transforming, not inventing.
  3. 3Ask a factual question about something obscure and specific to you — your street’s history, a small local company, a niche product’s exact release date. Then check the answer. This is where hallucinations show up.
  4. 4Ask: "What is today’s date and the latest news about X?" — see whether it searches the web or admits a cutoff. Different tools behave differently.
  5. 5Finally, paste a 20-row table from a spreadsheet and ask it to total a column. Check the arithmetic by hand. Many tools now run code to do this correctly — but only if they say so.
Prompt to paste
Summarise the text below in 3 bullet points for a busy reader. Then list anything in the text that seems like a claim I should double-check.

[paste text]

Check your work: You should come away with a feel for the pattern: transformation tasks are reliably good, obscure facts and precise numbers are where you must verify.

Developer walkthrough · Gemini · Claude · OpenAI

The whole product surface — chat apps, copilots, agents — sits on one API call: send text, get text back. Make that call once with each provider and you have seen 90% of the plumbing. Note the three inputs every SDK exposes: a model name, an instruction (system prompt), and the user message.

The one API call under everything. Install the SDK, set the API key as an environment variable, run.
python
# pip install anthropic          # export ANTHROPIC_API_KEY=...
from anthropic import Anthropic

client = Anthropic()

resp = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=300,
    system="You are a patient teacher. Avoid jargon.",
    messages=[{"role": "user",
               "content": "Explain what a context window is, in two sentences, to a project manager."}],
)
print(resp.content[0].text)
print(resp.usage.input_tokens, "tokens in,", resp.usage.output_tokens, "tokens out")
  1. 1Run it three times. Notice the wording changes — outputs are sampled, not deterministic. Set temperature to 0 and see it stabilise.
  2. 2Print the token counts. Paste a 2,000-word document into the message and watch the input count jump — this is the number you pay for.
  3. 3Ask it for a fact from last month. Observe the cutoff behaviour with no tools attached.

Practice

Everyone

Keep a one-week log: every time you use an AI assistant, write down the task and whether the answer was (a) useful as-is, (b) needed editing, or (c) wrong. By Friday you will know which tasks to trust it with.

Developers

Write a 30-line script that takes a text file, sends it to a model with the instruction "summarise in 5 bullets", and prints the summary and the cost (tokens × price). Run it on three different documents.

Pitfalls at this stage

  • •Treating the assistant as a search engine. It is a writer that may also be able to search — check whether it actually did.
  • •Pasting confidential data into a consumer tool without checking your organisation’s policy or the tool’s data settings.
  • •Believing citations. Models can invent plausible-looking references; open every link.

Free resources