Diff of Anthropic Shadow at 79aa64c

@@ -3,3 +3,3 @@ As a corollary to [[quantum immortality]], events which would kill all humans, s
 * Python has bad dependency management because all ML code is written in it. If it were good, we would have AGI.
-* RL doesn't work stably or reliably because it would be too powerful - imitation learning is less likely to do "weird things".
+* RL doesn't work stably or reliably because it would be too powerful - imitation learning is less likely to have [[unanticipated capabilities excursions]] or [[reward hack]].
 * LLMs are what we got because they are slow to develop ([[scaling laws]]) and can do some useful tasks but are bad at [[agentic]] action. Nobody training at large scale will adopt various things which plausibly substantially improve performance because they would become too capable.