Training: Guess → Check → Adjust
A model is billions of dials. Each wrong guess nudges them. Repeat trillions of times.
Prompt = today only.
The AI learned by playing a giant guessing game: guess the next word, check the answer, fix itself, billions of times. It does not keep learning from your chats.
Your prompt does not retrain the model. Anything it needs to know today, like company rules or a client's name, you tell it each time, or keep it in a Claude Project.
Hide a word: 💼 "approve the Q4 ___"
AI guesses "pizza" ❌
Right answer: "budget". Measure the miss.
Nudge the dials. Repeat for months.
Tokens: AI Reads in Puzzle Pieces
Short words = 1 piece. Long words get chopped: mis|under|stand|ing = 4.
The AI does not read whole words. It chops text into small puzzle pieces called tokens, and counts everything in pieces.
Wordy prompts and big pasted files use more pieces, so they cost more and hit limits sooner. Say it plainly and paste only what you need.
Brain Map: Similar Words Live Close Together
Bigger or fine-tuned models = sharper map = fewer mix-ups.
The AI keeps a map where words with similar meanings live next door. A small model sees that map like a blurry photo; a big model sees it in HD, so it spots small differences.
Add words that point to the meaning you want: 'bank' plus 'loan' lands in the money neighbourhood. When small differences matter, such as contracts, numbers or tone, use a bigger model.
Two spots on one cheek: a pimple (red, raised) and a mole (flat, brown). Click the brains above to change the photo.
Attention: The Spotlight on Your Words
The AI has one torch. Every word in your prompt shares its light.
The AI's attention gets shared around all your words. The more clutter you add, the easier it is for your main point to get lost.
Short and structured beats long and chatty. If the answer misses the point, trim the prompt before you add more to it.
Why this works: every extra word competes for the AI's attention. The more clutter, the easier it is for your main point to get lost. Tip: slide to Rambling to see tools 3 to 5 work.
Prediction: Super Autocomplete
Score every next word → roll a weighted die → add one word → repeat.
The AI writes like super autocomplete: it picks a likely next word, then the next, one at a time.
It sounds confident even when it is guessing. Ask it to say when it is unsure, and check facts and figures yourself.
Noise: Static in, Surprises out
Two kinds of noise: junk you paste in, and a little dice roll on every word.
Two kinds of noise: junk in your message, which it might answer instead of your question, and a little dice roll on every word, so the same question can get different answers.
Paste only what the task needs. For reports and facts, give exact wording and a fixed format; for ideas, ask for lots of options.
Chat apps roll a small dice on every word, so the same question rarely comes back word for word. You can't turn the dial in the Claude app, but precise wording leaves the dice fewer choices.
Memory Limit: Old Messages Fall Off
AI only sees the last 100 tokens. Add messages → your 🍕 and 💼 rules fall off the cliff.
The AI can only hold so much of a chat in its head. When it is full, the oldest messages fall off.
In long chats, repeat the key facts, or start a fresh chat with a short summary of what matters.
Drift, Pollution & Rot
Watch the 🧠 orb (AI's focus) get pulled off target.
In a long chat the topic slowly wanders, old mistakes stick around and things go stale, like a whiteboard nobody wipes.
One topic per chat. If answers start going odd, do not argue with it: start a new chat with a clean brief.
KV Cache: Don't Re-Read, Remember
AI saves a note for every past word, so each new word is 1 step, not a full re-read.
"Budget"
"$50,000"
The AI keeps notes on what it has already read, so it does not re-read your whole message for every new word it writes. That makes replies faster.
Some AI tools also reuse an identical start across messages. Keep the fixed part (instructions, policy document) at the top and the changing question at the bottom.
Latency is how long you wait before the AI types. KV Cache keeps latency flat and near-zero!
Trade-off: speed costs memory.