Reaction
The argument for improving AI safety through building and iteration, not delays
A post compares the approach to improving planes and cars by building them, spotting problems and making them safer.
TLDR
A post argues that building AI systems, finding problems and iterating will do more for safety than delays. It claims Anthropic has made distillation harder and cites Roon's forecast of progress in machine interpretability. The author argues that understanding an AI agent's inner workings would solve alignment.
Combined views
2.1K
3 Sources, first seen ago
