Control Feedback versus Evolutionary Feedback

AI Safety researchers often describe control as an ability to constrain goals pursued by AGI.

In control engineering, however, control involves a machine processing inputs (from sensors) into outputs (to actuators), such to constrain effects in the world.

In simplifying control, researchers can neglect actual feedback effects. Under evolution, code causes effects in world (as phenotypes), and effects in the world in turn cause changes to the code (as genotypes).

Can control feedback safely contain AGI evolutionary feedback? We argue that it cannot.

Some reading:

• "Lenses of Control" by Will Petillo https://alignmentforum.org/posts/NFYLjoa25QJJezL9f/lenses-of-control

• "The Control Problem: Unsolved or Unsolvable?" by Remmelt Ellen https://lesswrong.com/posts/xp6n2MG5vQkPpFEBH/the-control-problem-unsolved-or-unsolvable

• "Deconfusing AI and Evolution" by Remmelt Ellen https://lesswrong.com/posts/qvgEbZDcxwTSEBdwD/deconfusing-ai-and-evolution