What Are LLMs (Large Language Models)

The Simple Explanation

An LLM is a next-token prediction machine. Given some text, it predicts what comes next — like autocomplete, but trained on the entire internet and absurdly good at it.

Input:  "The capital of France is"
Model:  P(Paris) = 95%, P(Lyon) = 1.5%, P(a) = 0.8%, ...
Output: "Paris"

That's it. Every conversation, every code generation, every essay — it's all next-token prediction happening thousands of times in sequence.

---

How They're Built

Step 1: Architecture — The Transformer

The core innovation (Google, 2017 — "Attention Is All You Need" paper):

Input Text → Tokenize → Embeddings → [Transformer Layers × 96] → Probabilities → Next Token
                                          ↓
                                    Self-Attention:
                                    "Which words should I
                                     pay attention to when
                                     predicting the next one?"

Self-attention is the key mechanism: for each token, the model learns which OTHER tokens in the input are relevant. When processing "The cat sat on the ___", attention helps the model focus on "cat" and "sat" to predict "mat".

Step 2: Pre-Training

Feed the model trillions of tokens from the internet:

• Books, Wikipedia, code repositories, web pages, academic papers
• Training objective: predict the next token (self-supervised — no human labels needed)
• Cost: $10M-$100M+ in compute (thousands of GPUs for months)
• Result: a foundation model with broad knowledge

Step 3: Instruction Tuning (Fine-Tuning)

The raw pre-trained model just predicts text — it doesn't follow instructions well.

Fine-tune on instruction/response pairs:

Instruction: "Explain quantum computing in simple terms"
Response: "Quantum computing uses quantum bits (qubits) that can be..."

Now the model understands "when a human asks X, respond helpfully."

Step 4: RLHF (Reinforcement Learning from Human Feedback)

Humans rank model outputs from best to worst. Train a reward model from rankings. Use RL to make the LLM produce outputs the reward model scores highly.

This is how models become helpful, harmless, and honest — not just good at text prediction.

Pre-trained model → knows a lot but rambles
+ Instruction tuning → follows instructions
+ RLHF → actually helpful and safe
= Claude, GPT-4, etc.

---

What Parameters Mean

Parameters = the learned weights in the neural network. More parameters = more capacity to store patterns and knowledge. But also: more compute, more memory, more cost.

---

What LLMs Can and Can't Do

Can Do Well

• Generate fluent, coherent text
• Understand and follow complex instructions
• Write, debug, and explain code
• Translate between languages
• Summarize documents
• Reason through problems (with chain-of-thought)
• Roleplay, brainstorm, creative writing

Can't Do (Fundamental Limits)

The Hallucination Problem

You: "Who won the 2028 Olympics?"
LLM: "The 2028 Olympics were held in Los Angeles..." 
     (confidently generates plausible-sounding but potentially wrong details)

The model doesn't "know" things — it generates text that looks like it should follow your input. Sometimes what looks right IS right. Sometimes it isn't.

---

The Current Landscape (2025-2026)

---

Key Insight for Using AI Well

LLMs are tools for augmenting your thinking, not replacing it.

• They're best when you can verify the output
• They're dangerous when you trust blindly
• The better YOUR prompt, the better THEIR output
• Treat them like a very fast, very knowledgeable junior dev — verify everything

---

Resources

• 🎥 3Blue1Brown — But What Is a Neural Network?
• 🎥 3Blue1Brown — Attention in Transformers
• 📖 Andrej Karpathy — Intro to LLMs (1hr talk)
• 🔗 The Illustrated Transformer
• 📖 "Attention Is All You Need" paper (Google, 2017)

---

Next: 02 - Tokens & Tokenization