All videos
0:00 / 0:00
research

Understanding the inner thoughts of AI

Google DeepMind10 July 2026Watch on YouTube

Part of series

Ep. 7 · Prime Agent: Geheugen & Zelfcorrectie

Laat zien hoe Prime Agent via feedback-opslag en slash-refine functies zichzelf verbetert zonder handmatige hertraining.

View the series

What you'll learn

  • AI models don't contain fully formed thoughts, but billions of numbers forming internal representations from which patterns can be decoded
  • Mechanistic interpretability seeks to understand how AI systems work by making their internal structures visible, such as sparse autoencoders
  • Interpretability is essential for AI safety and alignment, as we must understand how AI makes decisions before systems become more powerful
  • Researchers use techniques like chain-of-thought monitoring to track how AI reasons and identify where problems might occur
  • There are inherent limits to what we can see inside AI systems, but the field is growing rapidly to make future AGI safer

Frequently asked questions

Why is understanding how AI 'thinks' important for AI safety?
Understanding AI's internal processes allows researchers to verify how systems make decisions and identify potential problems before AI systems become more powerful and critical.
What are sparse autoencoders and what do they tell us about AI?
Sparse autoencoders are elegant internal structures discovered in AI networks. They help reveal how AI systems organize and process information in ways that are more interpretable.
What are the limits of examining AI systems from the inside?
While mechanistic interpretability offers insights, there are inherent limits to what researchers can see and understand in AI systems, requiring further methodological innovation.

Topics

interpretabilitymechanistic interpretabilityai safetyalignmentdeepmind research

In this video

GoogleGoogle

Read next

Description from the channel

What if you were to peer inside the ‘mind’ of AI? You wouldn't find fully formed thoughts, just vast arrays of numbers. In this episode, Professor Hannah Fry is joined by Neel Nanda, to shine a light on an ongoing open area of research, interpretability. Neel and his team are trying to do something phenomenally difficult: understand an intelligence that didn't come with a manual. Together, they explore the cutting-edge "neuroscience" of artificial intelligence—revealing the surprising, elegant structures being discovered inside these networks (like spare auto encoders), the inherent limits of looking under the hood, and why interpretability is absolutely essential if we are to build safe, aligned and trustworthy AI as we move towards AGI. Learn more about this area of research via https://deepmind.google/ 00:00 Introduction 02:41 Motivation for interpretability research 04:01 Mechanistic interpretability 08:14 Chain of thought monitoring 18:14 Interpretability techniques 35:00 Auditing models for safety 48:53 What comes next for interpretability Intro visuals from Winston Duke for Visualising AI: https://winstonduke.com/Google-Deepmind-Visualising-AI ___ Subscribe to our channel https://www.youtube.com/@googledeepmind Find us on X https://x.com/GoogleDeepMind Follow us on Instagram https://instagram.com/googledeepmind Add us on Linkedin https://www.linkedin.com/company/deepmind/