Bottom line: all is not lost. Even if perfect containment is impossible, layered safeguards grounded in computation, information theory, and thermodynamics could likely give humanity time and tools to respond to even threats from a superintelligent AI adversary. (3/3)

Many assume a superintelligent AI could bypass any human-imposed limits. We challenge that assumption. Using standard security engineering approaches – threat modeling, protocols, primitives, and layered defenses – humans can impose real costs and barriers. (2/3)

Thanks! Depends on the type of risk you're talking about! For catastrophic risks, I think very long-term. There are plenty of other lesser risks that are short-term or already here though.

I also wrote a more accessible discussion of the report that was just published in @sciam.bsky.social called "Could AI Really Kill Off Humans?" Take a look!
www.scientificamerican.com/article/coul...

www.eff.org/deeplinks/20...