LLM Internals
Why Your LLM Prompt Costs More Tokens Than Expected
See how whitespace, punctuation, numbers, code and multilingual text shift token counts, with before/after examples and cheaper prompt formatting tips.
Read the guide
Guides
Deep dives between the course and the docs: each guide explains one decision you will actually have to make, with the trade-offs spelled out.
Training & Fine-tuning
RAG vs Fine-Tuning: Which One Should You Use for Your LLM?
RAG or fine-tuning? Use a simple decision framework to pick the right approach for knowledge, style and cost, with trade-offs and a worked LoRA example.
Read the guide
Inference & Cost
How Much VRAM Do You Need to Run an LLM Locally? Quantization Explained
How much VRAM does a local LLM need? Learn the weights-plus-KV-cache formula, what 4-bit quantization saves, and which GPU fits Llama 3 8B or 70B.
Read the guide
Fundamentals
Dot Product, L1 and L2 Norms: Linear Algebra for Machine Learning
Learn the dot product, L1 and L2 norms with worked examples, and see when ML uses each for loss, regularization and similarity. No heavy math required.
Read the guide
Inference & Cost
How to Price an AI Feature: Cost per User and Gross Margin
Learn how to price an AI feature: estimate LLM cost per request and per user, set usage limits, and protect gross margin with a simple worked example.
Read the guide
LLM Internals
Fine-Tuning Memory Math: Why Full Training Needs 16 Bytes per Parameter
Why full fine-tuning needs about 16 bytes per parameter: weights, gradients and Adam states. Compare it to LoRA and QLoRA with a worked 7B example.
Read the guide
LLM Internals
KV Cache Size Formula: Why Long Context Eats Your VRAM
Use the KV cache size formula to see why long context eats VRAM. Worked numbers for Llama 3 8B and 70B, plus ways to shrink the cache and serve more users.
Read the guide
Inference & Cost
Prompt Caching and Batch APIs: How to Cut LLM Costs
Prompt caching and batch APIs can cut LLM bills sharply. See how cached input and batch discounts work, with a worked cost example and checklist.
Read the guide
LLM Internals
Why Non-English Text Costs More Tokens: Turkish, Chinese, Arabic
Turkish, Chinese and Arabic often use more tokens than English for the same meaning. Learn why, how to measure the gap, and how it changes your API bill.
Read the guide
LLM Internals
How to Count Tokens in a Multi-Turn Chat Conversation
Learn how to count tokens across a full multi-turn chat, including per-message overhead and role tags, with a worked 3-message example and a budgeting checklist.
Read the guide
Inference & Cost
GPT-4o mini vs Claude Haiku vs Gemini Flash Cost
Compare GPT-4o mini, Claude Haiku and Gemini Flash on token pricing, context handling and real workload costs, with a practical calculator walkthrough.
Read the guide
Inference & Cost
DeepSeek API Pricing Explained: Is It Really Cheap?
A practical breakdown of DeepSeek API pricing, cache hits, and output costs, with a calculator walkthrough to see when it is truly cheaper.
Read the guide