How Anthropic Reads Claude's Thoughts

This introduces the interpretability tool the series calls a 'lens' — a method for reading a model's internal activity to see which concepts it's holding, before it speaks.

How it works

1. 1Find internal patterns that increase the probability of each possible next word.
2. 2Take a snapshot of the model's internal activity at a layer.
3. 3Measure how each activation influences the chance of every possible next token.
4. 4Read out the top concepts the model is 'thinking' about.
5. 5Use this to see hidden thoughts before the model outputs them.

Key points

• It reveals plans, risks and intentions early.
• It helps monitor alignment, safety and evaluations.
• It can catch issues (e.g. a hidden goal) before the model acts.
• It's an educational illustration of interpretability methods, not an exact product.

Why it matters

Reading internal activity turns a model from a black box into something researchers can observe and monitor — a foundation for safer, more trustworthy AI.

Frequently asked questions

What does the lens reveal?

The top concepts the model is internally focused on before it produces its next token.

Why is that useful?

It lets researchers detect risks or misalignment early, before the model outputs or acts.