How AI Actually Thinks — Visual Interactive Guide

Module 01 • How AI Reads Your Text

Tokens: AI Reads in Puzzle Pieces

Short words = 1 piece. Long words get chopped: mis|under|stand|ing = 4.

Human Words

AI Tokens Used

Long Words Chopped

Pick one:

Cut lines | 13 internal word cuts

AI Tokens Show IDs

👆 Tap a piece

💡 Why care?

You pay — and hit limits — by tokens, not words.

Module 02 • Embeddings

Brain Map: Similar Words Live Close Together

Bigger or fine-tuned models = sharper map = fewer mix-ups.

① Pick a brain

Pick one:

🧊

🔄 Auto-Rotate Show 3D Numbers

🖱️ Drag to rotate 🚀 Frontier 1T+ • "Bank" → 🌿 River & Nature

Map sharpness 🚀 Frontier Model (1T+)

Sharpness 98% (Ultra-Sharp)

Nearest to "Bank" 🌿 River & Nature

AI reply ✅ Understands Context

Module 03 • How AI Writes

Prediction: Super Autocomplete

Score every next word → roll a weighted die → add one word → repeat.

Pick one: Step 0 of 5

🌡️ Creativity (Temperature) 0.2 — Safe

0 Safe 0.7 1.5 Wild

Next-word odds

💡 A 25% word wins ~1 in 4 rolls — that's why answers vary.

Module 04 • Context Window

Memory Limit: Old Messages Fall Off

AI only sees the last 100 tokens. Add messages → your 🍕 and 💼 rules fall off the cliff.

Window 60 / 100 Tokens

🍕 SAFE

💼 SAFE

Pick one:

🕳️ Forgotten 0 tokens lost

🚫 Invisible

🪟 What AI sees (100)

✅ Both remembered

⚠️ Oldest 25t 50t 75t Newest ➔

💡 Ask below → add 💼 PDF twice → ask again.

🧠 What AI sees now

🎯 Test its memory

👆 Ask a question

Module 05 • Long Chats Go Bad

Drift, Pollution & Rot

Watch the 🧠 orb (AI's focus) get pulled off target.

🗺️ Where AI's focus goes

🧭 Drift

🧠 AI Focus Vector: (X: -0.55, Y: -0.45, Z: +0.60)

↔ X: Work vs Fun ↕ Y: Verified vs Polluted ↗ Z: Recent vs Buried

🔄 Drag to rotate 3D

💬 The chat Quality: 45%

🧭 Drift

Attention

Module 06 • Training

Training: Guess → Check → Adjust

A model is billions of dials. Each wrong guess nudges them. Repeat trillions of times.

🎓 Train vs. Prompt

Train = permanent.
Prompt = today only.

🎬 How a model learns

📚 1. Read the internet

🔄 Step 1,420 • 94%

1️⃣ Show

Hide a word: 💼 "approve the Q4 ___"

2️⃣ Guess

AI guesses "pizza" ❌

3️⃣ Score

Right answer: "budget". Measure the miss.

4️⃣ Adjust

Nudge the dials. Repeat for months.

Pick one:

Compare 3 setups

🧠 Brain & memory 🎓 5,000 examples

Brain dials 100% Tuned

🪟 Memory used up-front 12 / 100 Tokens (88% Free!)

Reply Speed

⚡ 0.3 sec

Cost Per Chat

💲 Very Low

Format Match

🎯 99% Exact

💬 You type

🤖 AI replies ✅ Right format

🎯 Fine-tune or not?

👆 Tap one

Module 07 • Speed Trick

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.

🔑 Key = label

"Budget"

📦 Value = meaning

"$50,000"

Pick one:

⚡ Saved notes → new word

⚡ 1 step

⏱️ Latency: 25ms

Writing word #6 of 10

Saved New Redone

🔍 Note for

🔑

📦

👆 Tap a word

📊 Score

This word 1 word computed

Total work 6 vs 21 ops

⏱️ Latency (Response Wait Time) 25ms (Instant)

0ms (Instant) 500ms 1,000ms+ (Lag)

Latency is how long you wait before the AI types. KV Cache keeps latency flat and near-zero!

💾 Memory used 6 Sticky Notes Saved

Trade-off: speed costs memory.

💡 Tip: keep the start of your prompt the same

👆 Tap one