Our Activities

The Mechanistic Interpretability Course is an intensive 1-month program focused on techniques for understanding the internal mechanisms of neural networks. This course combines theoretical learning with hands-on projects.

Mechanistic interpretability is a key area of AI safety research, aiming to make AI systems more transparent and understandable.

Curriculum Overview

• Foundations of neural network architectures
• Feature visualization techniques
• Attribution methods for understanding network decisions
• Advanced case studies from recent literature

Time Commitment

• 2 lectures per week (2 hours each)
• 1 practical session per week (3 hours)
• Individual project work (5-10 hours per week)
• Final project presentation

Prerequisites

• Strong programming skills (Python)
• Experience with deep learning frameworks (PyTorch preferred)
• Familiarity with basic neural network architectures
• Linear algebra and calculus

View Detailed Course Information

Our Weekly Discussion Group provides a casual forum for discussing recent papers, concepts, and developments in AI safety. These sessions are open to anyone interested in the field, regardless of prior knowledge.

Each week features a different topic, announced in advance through our mailing list and Telegram group.

Format

• 90-minute discussions led by a rotating facilitator
• Short presentation of the week's topic (15-20 minutes)
• Open discussion and Q&A
• Optional pre-reading materials shared in advance

Participation

No registration is required. Simply show up! If you're attending for the first time, we recommend arriving 10 minutes early to meet the organizers.