Session 7: Two Main Streams

Restructuring

• We reworked the activity pipeline for our weekly meetings, which goes as follows:
  • Self-research: Spend up to 1 ~ 1.5 hours researching the topics/working on the problems you want to discuss that day and then answer some related questions.
  • Discussion: Spend 15 ~ 20 minutes discussing what you just researched with someone who also researched the same thing/already has some foundations in it.
  • Presentation: Spend ~10 minutes presenting your findings to everyone. Then exchange questions about it as a way to think critically about your work.
• We split group activities into two main streams: Stream 1 and Stream 2. This is not a strict classification, as members are free to hop between these two streams whenever they want, as long as they try to follow our newly established activity pipeline.

Two Main Streams

• In the 1st stream, all of us built a better understanding of some of the foundational techniques and architectures used in modern AI systems by:

  • Doing some basic research on neural networks and machine learning
  • Going through some questions on these topics in a few minutes
  • Discussing the implications of these answers for AI development, and sometimes, AI Safety

• For everyone interested in Mechanistic Interpretability, activities were:

  • Try to read and research further on the topics presented in one or more of the following articles/papers:
    • Zoom In: An Introduction to Circuits
    • Toy Models of Superposition
    • Towards Monosemanticity
    • Interpretability in the Wild
  • Then, attempt to:
    • Summarize what the central claim is. Express some uncertainty you may have.
    • Find evidence or arguments that support or reject the claim.
    • Answer some questions related to the reading.

• In the 2nd stream, we:

  • Read through the course on AI policy from BluedotImpact.
  • Researched and tried to answer some questions related to recent developments in AI to get an idea of the landscape, mainly to see how fast it’s moving
  • After this, we each chose one of the following questions, spent 15 minutes researching and writing up a 1-page report on it. Then we presented what we could research to everyone and answered everyone’s questions:
    • What is the EU AI safety act about? How does this help prevent AI risks?
    • What is the current AI safety paradigm in developing countries?
    • What is Moore’s law? Does this still hold up to today? What does this mean for AI development growth?
    • Has there been any examples of AI misuse? What about accidents? What about misalignment? Distill and demonstrate them.