GDM Mech Interp Team Updates on Sparse Autoencoders and Interpretability Research

The Google DeepMind mechanistic interpretability team has released a progress update, detailing their recent advancements in scaling sparse autoencoders (SAEs) to larger models and conducting foundational research on SAEs. This update is inspired by the Anthropic team's excellent monthly reports and aims to share insights that are valuable to the broader AI community, particularly those working on mechanistic interpretability.

Key Takeaways

• Scaling Sparse Autoencoders: The team has made progress in scaling SAEs to larger models.
• Basic Science on SAEs: Initial steps toward understanding fundamental aspects of SAEs.
• Infrastructure Improvements: Enhancements that could benefit a wider range of mechanistic interpretability researchers.

Scaling Sparse Autoencoders

One of the primary goals of the GDM Mech Interp team is to scale sparse autoencoders (SAEs) to larger models. Here are some key points:

• Preliminary Results: The team has achieved promising results in scaling SAEs, though they are still evaluating the full impact.
• Steering Vectors: They have improved and interpreted steering vectors, which are crucial for controlling the behavior of neural networks.
  • Improvements: Enhanced stability and interpretability of steering vectors.
  • Interpretation: Better understanding of how these vectors influence model outputs.
• Ghost Grads: The team has also made progress in improving ghost gradients (ghost grads), a technique used to approximate gradients in sparse models.
  • Enhancements: More accurate and efficient ghost grad calculations.
  • Impact: Improved training dynamics and performance.

Basic Science on SAEs

The team is conducting foundational research to better understand the behavior of SAEs. This includes:

• Initial Steps: The team has taken initial steps in several areas, such as:
  • Sparse Approximation Algorithms: Replacing SAE encoders with inference-time sparse approximation algorithms.
    • Benefits: Potential for faster inference and reduced computational overhead.
  • Model Interpretability: Exploring new methods to interpret the internal mechanisms of SAEs.
    • Techniques: Using visualization tools and feature attribution methods to gain insights into model behavior.

Infrastructure Enhancements

The team has also made infrastructure improvements that could benefit a broader range of mechanistic interpretability researchers:

• Modular Codebase: Developed a more modular codebase to facilitate easier experimentation and collaboration.
  • Benefits: Simplified workflow for researchers working on various aspects of SAEs.
• Benchmarking Tools: Introduced new benchmarking tools to evaluate the performance of different SAE configurations.
  • Impact: Helps in identifying optimal settings and potential bottlenecks.

Confidence and Evidence

The team has been transparent about their level of confidence in the results and the evidence supporting their conclusions. They encourage the community to critically evaluate these findings and provide feedback.

• Confidence Levels: Clearly stated for each result, ranging from initial promising steps to more robust findings.
• Evidence: Detailed descriptions of experiments, data, and methodologies used to reach their conclusions.

Conclusion

The GDM Mech Interp team's progress update provides valuable insights into the current state of sparse autoencoder research. Their work on scaling SAEs, conducting foundational science, and improving infrastructure is expected to benefit not only those working with SAEs but also the broader AI community interested in mechanistic interpretability.