AI Fundamentals

How language models actually work — tokens, embeddings, context windows — explained with demos you can run and change.

• 100% Accurate and Useless: Pick a Model Before You See Its ScoreNobody overfits on purpose. They overfit because the number in front of them is going up. Tune a classifier on training data here, commit to it, and only then find out what it scores on data it has never seen.✓ numbers produced by the code on this page
• 99% Accurate, and Mostly Wrong: Why Rare Things Break Good DetectorsA test that gets 99% of its calls right can still be wrong about 92 of every 100 people it flags. Two dials here - how good the test is, and who you point it at - and only one of them can fix that. Find out which before the page tells you.✓ numbers produced by the code on this page
• Your Prompt Has 3,840 Right Answers, and It Picked OneWrite a post about our launch is six words and not a request for one thing. Count the decisions it leaves open and the options multiply out to 3,840 different replies, every one of them a fair reading. Send it here, watch the model fill in your blanks, and see what specifying actually buys.✓ numbers produced by the code on this page
• Trim or Summarise: The Cheap Fix That Makes Your Agent Forget Why It Is HereAn agent's context grows until something stops it, and there are two ways to stop it. On the same schedule they cost within 60 tokens of each other, and one of them ends the run having lost every fact the job depended on. Run twenty turns here and choose.✓ numbers produced by the code on this page
• What to Hand an AI and What to Keep: It Is Not About DifficultyDelegating everything to an AI and checking it saves barely 13% of a working day. Choosing task by task saves 28%. Run ten jobs here and find the axis that decides - not how hard the work is, but how long it takes to check somebody else's answer.✓ numbers produced by the code on this page
• Why Your Claude Code Prompt Is Not a GuardrailInstructions in a prompt or CLAUDE.md are requests the agent can complete its way around. Deny rules, PreToolUse hooks and files the agent cannot write are enforced outside its turn. Here is the difference, and what we run in production.✓ sourced to official docs
• Grok Bot: What xAI Documents, and What Is Already Wrong OnlinexAI and Cursor shipped Grok Bot in beta on 11 August 2026. Here is what the official documentation actually says, including the shared-computer detail that most coverage has backwards, and why it changes the security question.✓ sourced to official docs
• How Embeddings Work, Counted by HandAn embedding turns a word into numbers so that similar words land near each other. Build real word vectors from your own text, see why cat and dog score alike without ever appearing together, and why similarity is measured by angle rather than distance.✓ sourced to official docs
• What Is a Token? Train a Tokenizer and Watch One FormLanguage models read tokens, not letters or words. Train a real byte-pair encoding tokenizer on your own text, watch the vocabulary build itself merge by merge, and see why models struggle to spell.✓ sourced to official docs
• Temperature and Top-P: The Two Dials, and Why They FightTemperature reshapes the probability distribution over the next token; top-p cuts off its tail. Drag both and watch the arithmetic — including why changing one silently changes what the other does.✓ sourced to official docs
• AGI: One Word, Two Definitions That Do Not AgreeThe argument about whether AGI is close never resolves, and the reason is not the evidence. The word names at least three separate axes, and the major definitions draw the line across different ones. Draw your own and see which you picked.✓ sourced to official docs
• Why ChatGPT Forgets: It Never RememberedA chat model is stateless. Every turn, the app re-sends the entire conversation — which is why long chats start forgetting the beginning, and why they cost far more than the number of words you typed.✓ sourced to official docs
• How Many R's in Strawberry? Why That Question Is UnfairA model that writes working code will confidently miscount the letters in a word. It is not a reasoning failure — the letters were destroyed before the model saw anything. Train a tokenizer and watch it happen.✓ sourced to official docs
• Prompt Injection Is Not a Phrase, and a Blocklist Scores Worse Than GuessingSort twelve inputs into attacks and ordinary work, then watch a keyword filter score four out of twelve on the same set - missing four real attacks and blocking four harmless sentences whose only fault is discussing security.✓ numbers produced by the code on this page