What Matters Right Now In Mechanistic Interpretability?

This is a talk I gave to my MATS 9.0 training scholars about the big picture of mech interp - as of Oct 2025, what had changed? What are the most important things to work on? What are common mistakes? How do I recommend approaching the field? Notes: https://docs.google.com/document/d/1d... 00:00:00 What Matters Now in Mech Interp 00:02:16 The Rise of Reasoning Models 00:04:49 Q&A: Architecture Changes 00:08:58 Research Priorities 00:10:13 AI Psychology & Misalignment 00:12:05 Debugging Model Failures 00:14:53 Understanding vs. Control 00:24:07 Case Study: Golden Gate Claude 00:30:00 A Pragmatic Research Philosophy 00:35:03 Pragmatism, Baselines, and Simplicity 00:46:12 Diagnosis vs. Solution 00:53:55 Do We Need Precision?

Join Today
What Happened With Sparse Autoencoders?
▶︎

What Happened With Sparse Autoencoders?

How Will Mech Interp Help Make AGI Safe?
▶︎

How Will Mech Interp Help Make AGI Safe?

Mechanistic Interpretability explained | Chris Olah and Lex Fridman
▶︎

Mechanistic Interpretability explained | Chris Olah and Lex Fridman

Interpretability: Understanding how AI models think
▶︎

Interpretability: Understanding how AI models think

The Bitter Lesson for Biology — Adam Green on Virtual Cells and Scaling Laws
▶︎

The Bitter Lesson for Biology — Adam Green on Virtual Cells and Scaling Laws

Mechanistic Interpretability and How LLMs Understand
▶︎

Mechanistic Interpretability and How LLMs Understand

The Story of Mech Interp
▶︎

The Story of Mech Interp

Mechanistic Interpretability for NLP: One-stop Guide for Everything you Need to Know
▶︎

Mechanistic Interpretability for NLP: One-stop Guide for Everything you Need to Know

How Reasoning Models Break Mechanistic Interpretability Techniques
▶︎

How Reasoning Models Break Mechanistic Interpretability Techniques

How AI Learned to Teach Itself [JEPA]
▶︎

How AI Learned to Teach Itself [JEPA]

Reinventing Entropy | Compression is Intelligence Part 1
▶︎

Reinventing Entropy | Compression is Intelligence Part 1

The French Do Not Care About Work
▶︎

The French Do Not Care About Work

The Dark Matter of AI [Mechanistic Interpretability]
▶︎

The Dark Matter of AI [Mechanistic Interpretability]

Scaling interpretability
▶︎

Scaling interpretability

Science of Misalignment
▶︎

Science of Misalignment

Margin Call - "Sell it all. Today." 👆🤘👆
▶︎

Margin Call - "Sell it all. Today." 👆🤘👆

Why you should care about AI interpretability - Mark Bissell, Goodfire AI
▶︎

Why you should care about AI interpretability - Mark Bissell, Goodfire AI

Yann LeCun's $1B Bet Against LLMs [Part 1]
▶︎

Yann LeCun's $1B Bet Against LLMs [Part 1]

If Prime Numbers Become Increasingly Rare, Then Why Do They Keep Showing Up In Pairs?
▶︎

If Prime Numbers Become Increasingly Rare, Then Why Do They Keep Showing Up In Pairs?

Ilya Sutskever – We're moving from the age of scaling to the age of research
▶︎

Ilya Sutskever – We're moving from the age of scaling to the age of research