Rather than try to train a model to be better, assuming it ‘thinks’ like us.

**Mechanistic interpretability** is a field that seeks to understand the internal reasoning process of AI, so that we can the _how_ and _why_ models produce certain outputs.

Frontier models contain billions of parameters spread across hundreds of layers. The core challenge is that it’s hard to understand what each part does.


How does AI think?

Several techniques help us make sense of all this complexity, like:

- Probing classifiers
- Feature steering
- Chain of thought monitoring

With this knowledge, we could tweak the model's internals to make it ‘unlearn’ certain things or get it to think differently.

But knowledge is dual-use. You can’t just get an AI to unlearn how to make bioweapons because similar knowledge is useful for making vaccines. Neither can you get an AI to unlearn ‘lying’ because truth is a very vague concept. Doing this sometimes makes the model worse.


Making sense of model internals

This is the tradeoff that AI companies make. How much safer are we willing to make a model at the expense of capabilities? How long are they willing to do this and delay the next release?


Capable vs safe

Technical AI Safety

What's actually happening inside these black boxes?


Understanding AI

Why is building safe AI so technically difficult?


The technical challenge with AI

How do we teach AI to behave the way we want?


Training safer models

How do we know if the AI is actually safe?


Detecting danger

What happens when training and detection fail?


Minimising harm

How can you contribute to making AI safer?


Technical AI Safety

How does AI think?

Resources (1 hr 10 mins)