AXRP - the AI X-risk Research Podcast 38.2 - Jesse Hoogland on Singular Learning Theory
Nov 27, 2024
Jesse Hoogland, executive director of Timaeus and researcher in singular learning theory (SLT), shares fascinating insights on AI alignment. He dives into the concept of the refined local learning coefficient (LLC) and its role in uncovering new circuits in language models. The conversation also touches on the challenges of interpretability and model complexity. Hoogland emphasizes the importance of outreach efforts in disseminating research and fostering interdisciplinary collaboration to enhance understanding of AI safety.
AI Snips
Chapters
Transcript
Episode notes
SLT Connects Geometry To Generalization
- Singular learning theory (SLT) links loss-landscape geometry to why neural nets generalize so well.
- Timaeus uses SLT to build tools for evaluation and interpretability relevant to safety.
Timaeus Balances Theory And Applied Experiments
- Jesse describes Timaeus as a research org focused on applying SLT to practical problems by training models and running learning-coefficient sweeps.
- He contrasts theoretical work (Daniel Murfitt) with Timaeus's applied experiments and outreach.
Do Active Outreach To Scale SLT Impact
- Do outreach by giving talks and doing personalized engagement to get others familiar with SLT tools.
- Use these contacts to later hand off techniques to labs, policymakers, and other decision-makers.