I really enjoyed @yonashav's paper on how we might create a world where all large training runs can be monitored - feels like the best AI governance proposal I can recall seeing! Makes me optimistic there are policies simple enough to be realistic, but useful enough to matter