Inside Claude's Brain

This explores what happens inside a model before it speaks — the idea that most of the model's reasoning happens silently, in an internal 'workspace', before any output token appears. It's an accessible framing of AI interpretability research.

How it works

1. 1A user prompt comes in (e.g. 'book a trip to Paris under $800').
2. 2The tokenizer turns text into tokens; transformer layers pass information through millions of activations.
3. 3In the internal workspace, the model holds silent thoughts, concepts and plans.
4. 4It reasons and decides what to do next (plan steps, use tools, check constraints).
5. 5Finally it generates and streams the response as output tokens.

Key points

• The traditional view (prompt → model → answer) is too simple; there's hidden thinking in between.
• Understanding the internal workspace unlocks better control, safety and monitoring.
• Observe → understand → improve → control.
• This is an educational take on interpretability research, not an official spec.

Why it matters

Interpretability — understanding a model's internal reasoning — is key to building safer, more controllable AI. This series makes those research ideas approachable.

Frequently asked questions

Is this how models literally work?

It's a simplified, educational framing of interpretability research about a model's internal reasoning before it outputs text.

Why care about internal reasoning?

Seeing hidden thinking helps with safety, debugging and steering models more reliably.