The Adolescence of Technology

In general, he gave quite a few realistic-seeming examples of how AI can be used to accomplish a lot of bad things. If I were a terrorist, I’d read this as a script.

Section 1

• AI will become like a country of geniuses in a data center In the next couple of years
• Even though there is a small chance that AI can be used for catastrophically bad means or become catastrophically bad itself, we should still plan for it.
• We should be careful with our legislation, because technology evolves faster than we can adapt.
  • He proposes chip export control and transparency on risks.
• Somewhat confident that alignment training (e.g. Constitutional AI) plus mechanistic interpretability is the way to go.

Section 2

• terrorists or mentally ill people can use AI to help them create weapons of mass destruction
• he’s most concerned about bioweapons because they have the most potential to destroy the entire human race easily
  • gives a decent amount of examples of these
• google and anthropic have both implemented classifiers on their models to check for this kind of thing

Section 3

• State actors can do a lot of bad things with AI. Surveillance state, powerful offensive capabilities for which we don’t know if defense will work properly, etc.
• CCP is the most worrying. We should stop exporting chips to them, it’s like exporting uranium for nuclear weapons