Under the Hood

Before you build more, it helps to feel what's actually happening when you talk to an AI.

Every language model starts as a completion engine — it predicts what text comes next, like autocomplete on steroids. That's it. No understanding, no intent, no helpfulness. Just: “given this text so far, what word is statistically likely to follow?”

The AI you're talking to in Claude Code went through additional training (RLHF) that taught it to interpret your text as a request and respond like an assistant. That's a learned behavior, not a fundamental capability.

Try it yourself. Type the same text into both models and watch the difference: