New member initiation
We initiated our newer members - Welsh and Kappie - on the basics of AI and AI alignment.
Self-research
We spent the first half of the meeting (2pm - 3:45pm) researching and taking notes on topics that we found interesting. Generally, the group was split into two streams, as per our established activity pipeline: Stream 1 and Stream 2.
Here’s a list of what each of us did research on:
- Welsh: AI watermarking
- Mogu: Mechanistic interpretability (overarching high level methodology)
- Gump: Mechanistic interpretability (features, circuits, universality)
- Kappie: Data input control
- Qronox: Developmental interpretability
- Skrubz: Economic potential of AI
Presentation
- In the latter half of the meeting, each of us went up and presented what we learned from the research session.
- During or after each of these presentations, we would ask and answer questions for clarification and criticism. We learnt a lot from these presentations!
overview tree of devinterp by qronox