Jacob Kelter
@jzkelter.bsky.social
Director at NetLogo Center @netlogo.bsky.social
Here's my short summary: bsky.app/profile/jzke...
My understanding of "If Anyone Builds It, Everyone Dies":
1. AI built on current techniques will have alien values
2. A superintelligence will be able to steer the world to realize its values
3. By default, alien values being maximized → humanity dies
4. Therefore, IABIED (book's title)
1. AI built on current techniques will have alien values
2. A superintelligence will be able to steer the world to realize its values
3. By default, alien values being maximized → humanity dies
4. Therefore, IABIED (book's title)
September 21, 2025 at 5:52 PM
Here's my short summary: bsky.app/profile/jzke...
A key point of the book: even if you disagree with (1), it's irresponsible to race ahead with AI unless you can confidently show that AI *won't* have alien values. If it *might* have human-aligned values, that doesn't justify racing ahead.
September 21, 2025 at 4:03 PM
A key point of the book: even if you disagree with (1), it's irresponsible to race ahead with AI unless you can confidently show that AI *won't* have alien values. If it *might* have human-aligned values, that doesn't justify racing ahead.
I think the only one of these that can plausibly be argued with is (1), and that is where most of the reasoned debate is happening.
September 21, 2025 at 4:03 PM
I think the only one of these that can plausibly be argued with is (1), and that is where most of the reasoned debate is happening.
Very cool model! I'm implementing a #NetLogo version, and I have a question: is there a real distinction between h_in vs h_out and γ_in vs γ_out, or could there just be 4 arbitrary traits? Since any link can be harmful or helpful based on harm/help ratio, the traits seem arbitrary.
August 15, 2025 at 5:38 AM
Very cool model! I'm implementing a #NetLogo version, and I have a question: is there a real distinction between h_in vs h_out and γ_in vs γ_out, or could there just be 4 arbitrary traits? Since any link can be harmful or helpful based on harm/help ratio, the traits seem arbitrary.