Neural Ninjas · How LLMs Actually Work
LLMs Don't Understand — They Tokenize and Compute
Same sentence, two realities. Toggle, tap, and drag to see which one is true.
💬 "Your LLM doesn't understand your question. It computes a statistically plausible answer. The magic isn't consciousness — it's math at scale."
Same sentence. Completely different reality.
Tap a word for what the model actually sees
Tap any word above to see what the model actually stores for it.
| What beginners think | What actually happens |
|---|
| 🧠 "It understands my question" | 📊 It maps tokens to training patterns |
| 📖 "It reads and comprehends" | 🔢 It processes numerical vectors |
| 💬 "It has reasoning" | 🎲 It predicts the most probable next token |
Watch a sentence stop being words
Tokenize → Embed → Process → Decode
At no point did the model "understand" — it just computed.
What this showsPress "Run pipeline" and watch words stop being words.
It's not thinking — it's sampling
Drag temperature, then sample the next token
"The cat sat on the ___"
What this showsLow temperature always picks "mat." High temperature spreads the odds — the model isn't more "creative," the sampling is just less certain.