What is an LLM?
LLM stands for Large Language Model. It's the technology behind ChatGPT, Claude, Gemini, and every other AI you've talked to. At its simplest: an LLM is a computer program that predicts what word comes next.
The model generates one word at a time. Each new word joins the prompt, and the model predicts the next one.
A neural network that learned patterns from reading the internet
An LLM is a neural network trained on a massive amount of text - books, articles, code, conversations, websites. During training, it learned patterns: grammar, facts, reasoning, writing styles.
When you give it a prompt, it doesn't "think" or "know" things the way a person does. It predicts the most probable next word based on everything it read during training.
Do that repeatedly - predict the next word, add it to the text, predict the word after that - and you get coherent paragraphs. The model was never taught to write essays. It just got very good at guessing what word comes next.
Write, answer, reason, and follow instructions
- Write text - emails, articles, code, stories, poems
- Answer questions - summarize documents, explain concepts, translate languages
- Reason through problems - solve math, debug code, plan projects
- Follow instructions - "rewrite this in a professional tone," "make this shorter"
- Know recent events - trained on data up to a cutoff date
- Access the internet - unless connected to a search tool
- Be factually reliable - it can hallucinate: confidently state incorrect information
Three steps: train, fine-tune, generate
Training
Read billions of words from the internet and books. Learn patterns - not memorize facts. The model discovers that "cat" often follows "the," and that paragraphs about physics use certain words together.
Fine-tuning
Practice on curated examples with human feedback. Learn to be helpful, accurate, and safe. This is where the model learns to say "I don't know" instead of making things up.
Inference
When you type a prompt, the model generates one word at a time. Each word is influenced by all previous words. This is the step you experience when chatting.
The key mechanism is called a Transformer architecture. It lets the model look at every word in your prompt simultaneously, rather than reading left to right like a person would.
Start free, switch when you hit limits
Strong all-around, good for creative writing and broad reasoning.
Strong at coding, long documents, and careful reasoning.
Strong at multimodal (images + text), integrated with Google tools.
Open-source, 10-34x cheaper, strong at coding and math.
You can also run local models (Llama, Qwen, via Ollama) on your own computer. They're free and private, but less capable than cloud models. Start with whatever is free and convenient.
Four things every beginner should understand
They hallucinate
LLMs can and do make things up, confidently. Always verify important facts. If an answer matters, double-check it.
They have knowledge cutoffs
Trained on data up to a specific date. They don't know what happened after. Ask when the model's training data ends if recent events matter.
They have context limits
Can only "remember" a certain amount of your conversation at once. Very long chats may lose earlier parts.
They are tools, not oracles
Best used for drafting, brainstorming, summarizing, and coding - not for final decisions without human review.
"An LLM doesn't know things. It predicts words. The fact that prediction looks so much like intelligence is the surprising part."
LLMs predict the next word - everything else (essays, code, conversation) emerges from that simple mechanism.
Verify important outputs. LLMs are tools, not sources of truth. Trust but verify.