You've used the machine for eight levels. Now you get to see inside it. Not the maths — the five ideas that explain the weird stuff you've probably already noticed: why long chats drift, why it can't say who won last night, why some models cost 100× others.
Training is the studying: months of finding patterns in enormous amounts of text. The model is what's left after — billions of numbers encoding those patterns, frozen. When you chat, it isn't learning. It's applying.
This explains two things you've met already. Why the knowledge cutoff exists (Level 01): the studying stopped on a date; anything after isn't in the numbers — unless the app bolts on web search. And why it doesn't remember you between chats: talking to it doesn't change it. Nothing you type rewires the model. Each conversation starts from the same frozen snapshot.
The model's only "memory" of your conversation is the conversation itself — re-read in full, every single reply. The amount it can re-read at once is the context window. Long chats drift because the window fills — and the early stuff effectively falls off the table.
You've felt this: forty messages in, it forgets what you said at the start, contradicts itself, gets vaguer. Now you know why — and the fix is mechanical, not magical:
When a long chat gets fuzzy: start a new chat and re-feed only what matters — "Here's where we got to: [3-line summary + the current draft]. Continue from here." Fresh window, zero clutter, all signal. This one habit fixes 90% of "the AI got dumb" complaints.
Tokens are how it reads: text chopped into chunks of a few characters. "Understanding" might be one token; "misunderstanding" three. This explains the small weirdness — counting letters in a word is genuinely hard when you can't see letters — and it's the billing unit: every token in and out is compute, and compute is money.
Model tiers follow directly: small models (fewer numbers) are fast and cheap and fine for everyday asks; big models are slower, pricier, and visibly better at hard reasoning. The free tier of most apps runs the smaller ones — which is why this course never needed the paid tier, and why the paid tier exists.
Everyday questions, summaries, drafts → any model is fine. Hard reasoning, long documents, code, anything where wrong is expensive → use the best model you can get, and give it room to think ("think step by step before answering" genuinely helps). Matching the model to the job is a builder's habit — most people never learn it.
Training made the numbers; chatting doesn't change them. Cutoffs and blank slates both follow.
Long chats drift because it fills. Fresh chat + re-feed the essentials = sharp again.
Cheap-fast for everyday, deep-smart for hard. Choosing on purpose is a builder's habit.
← back to all 10 levels · stuck on anything? ask peter.