Introduction to LLMs

Introduction

An LLM is a transformer trained on huge text corpora to predict the next token. It is the brain of every modern agent. Picking the right model is the first architectural choice — it shapes cost, latency and capability.

Beginner analogy: LLMs are like cars — you don't build one, you pick the right model for the road (use case): a Ferrari (GPT-4, Claude Opus) for hard reasoning, a Honda Civic (GPT-4o-mini, Haiku) for daily volume.

Understanding the topic

Core concepts:

• Closed models: OpenAI GPT-4/o-series, Anthropic Claude, Google Gemini.
• Open models: Meta Llama, Mistral, Qwen, DeepSeek.
• Reasoning models: OpenAI o1/o3, Claude 3.7 thinking, DeepSeek R1.
• Trade-offs: cost · latency · context · multimodality · privacy.
• Most agent systems use 2-3 different models for different steps.

Syntax reference

Visual workflow / architecture:

┌──────────────┐
│  Frontier    │  GPT-4o, Claude 3.5 Sonnet
└──────┬───────┘
       │
┌──────▼───────┐
│  Mini        │  GPT-4o-mini, Haiku
└──────┬───────┘
       │
┌──────▼───────┐
│  Open        │  Llama 3, Mistral, Qwen
└──────────────┘

Real-world use

ChatGPT defaults to GPT-4o; Cursor uses Claude 3.5 Sonnet; Perplexity routes across many models; most startups mix Sonnet + Haiku.

Best practices

• Benchmark on YOUR task — leaderboards lie.
• Use the cheapest model that passes evals; upgrade only on the steps that fail.

Common mistakes

• Picking only one model — you waste money on easy steps and quality on hard ones.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Difference between closed and open-source LLMs?
• Q2. Why mix models in one agent?
• Q3. When pick a reasoning model?
• Q4. Scenario: your monthly LLM bill is $10K. How do you cut 60%?