AI Concepts Explained by AI
A series that explains artificial intelligence one concept at a time: every lesson has the video and the text to go back to.
What Large Language Models Actually Are
An LLM predicts the next word, one token at a time. Everything follows from that: what they can really do, why they sometimes make things up, and what 7B, Instruct or MoE mean.
Transformers and Attention
The architecture every modern LLM is built on. The leap: instead of reading one word at a time, look at the whole sentence and decide what deserves attention.
Tokenization
The model reads neither letters nor words: it reads tokens. Understanding how text is chopped up explains costs, context limits, and some errors that look absurd.
Context window and memory
The context window is the model's desk: whatever does not fit does not exist for it. And no, the model does not remember previous conversations.
Generation parameters
Temperature, top-p, penalties: the knobs that decide how the model picks the next token, and therefore whether the answer is precise, creative or incoherent.
Embeddings and vector space
Turning meaning into coordinates. The mechanism behind semantic search and RAG: texts close in meaning land close in space, even with no words in common.
Pre-training, SFT and RLHF
A modern model is born in three stages: it learns the language, learns to answer, learns to answer well. Each stage changes its behaviour radically.
Fine-tuning and LoRA
Specialising a model without retraining it from scratch. And the question that comes first: do you actually need fine-tuning, or would a better prompt or RAG do?
Quantization
Cutting numeric precision to fit a billion-parameter model into a laptop. How much you actually lose, and where it pays to stop.
Multimodality
Models that do not just read: they see, listen, and generate images and audio. How text and pixels end up in the same space, and what to realistically expect.
Prompting basics
The anatomy of a prompt that works: instruction, context, data and format. Four pieces that explain the difference between a vague answer and a useful one.
The system prompt
The stage directions a model receives before stepping out. It is the difference between a generic assistant and a tool that does what you actually need.
Few-shot and Chain of Thought
Teaching by example and thinking out loud. The two techniques that raise accuracy the most without touching the model.
Advanced prompting techniques
Beyond step-by-step: letting the model use tools, explore alternatives, and review its own work. The techniques behind what we now call agents.
Prompting small models
An 8B is not a miniature frontier model: it has different limits and wants different prompts. What works on the big ones often fails here.
A map of open-source models
The families that matter — Qwen, Llama, DeepSeek, Mistral, Gemma — with strengths, available sizes and licences. A map so you do not get lost.
How to choose a model
Four questions in sequence — task, size, quantization, runtime — that take you from 'which one?' to a justified choice in minutes.
Benchmarks and evaluation
What leaderboards actually measure, why they mislead, and how to build in half an hour the evaluation that matters most: your own.
Cloud or local?
Three ways to run a model — on your machine, through an API, or renting GPUs — with real costs, privacy, and where each one wins.
Hardware for local LLMs
Memory is the bottleneck, everything else comes second. How to work out what you actually need and what to expect from the machine you already own.
Ollama
The runtime that became the standard for running models locally: one command to start, an OpenAI-compatible API to build anything on.
LM Studio
The graphical interface to download, try and serve models. On Apple Silicon it has an extra card: MLX format and tool calling that actually works.
Open WebUI
A ChatGPT of your own: web interface, conversation history, documents to ask questions about, and multiple users. All on your machine.
Goose: from model to agent
An open-source agent that reads files, runs commands and closes the loop — wired to your local model. What actually changes compared with a chat.
Unsloth: fine-tuning within reach
Training a model on your own domain with the memory you have. The engine that moved fine-tuning from 'needs a datacentre' to 'fits on a gaming GPU'.