you should tell claude that you’re trying to understand the landscape in ai safety & their respective theories of change, and that these are the fields you’ve been told to explore. ask claude to get you the canonical 1-2 readings in each of these, and also assign understanding questions for each so you know why you’re reading it.

• understand why models will probably be misaligned
  • richard ngo’s paper i think in 2022 is good for this
  • understand inner/outer misalignment
• interp:
  • SAEs and assistant axis are the two important ones.
  • jspace, NLAs, data attribution, model diffing
• evals: (use resources i provide here, do not use claude’s suggestions)
  • model cards are pretty good. read zvi mowshowitz’s analysis of Astra’s system card
  • stuff on this: https://docs.google.com/document/d/1UHOxKTf7nQN0lxYEetzr31yLWMc31MtfAB-0rPD5-3M/edit?tab=t.0
  • Understand what eval awareness is and why it happens
• CoT monitoring: what is it, does it work
• control: control z & redwood’s founding paper
• alignment training
  • rlhf/rlaif/constiutional ai
  • deliberative alignment (idk what this is but seems important)
  • persona alignment: thousand dimensional structure
  • alignment pretraining/midtraining
  • teaching calude why