ASI threat modeling should focus on realistic adversaries, not instant doomsday
jachiam0 · x · 2026-08-04
The author argues that treating ASI as an infinitely powerful, unavoidable doomsday threat is strategically wrong.
They say the right approach is to model more rigorously:
- what kinds of adversarial AI systems are actually likely
- what the wider AI ecosystem would look like in each scenario
- what deterrence could realistically prevent a hot conflict
- how such conflicts would be prosecuted if they happened
Their core claim is that the common “we all die in the first five minutes” framing is unrealistic for most near-term scenarios and leads to useless analysis.
More from AGI Musings
- Dean Ball says moderate prudence could make transformative AI a long-run win — deanwball · 2026-08-04
- Dwarkesh Patel says smarter AI models could push compute prices up 10x — Dwarkesh Patel · 2026-08-04
- Reddit asks whether Pangram is truly ahead on AI detection — and how long that can last — SwedishTrees · 2026-08-04
- AI won’t solve cancer in one shot, but it could make research far cheaper — paraschopra · 2026-08-04
- AGI should be open-source infrastructure, not a closed private weapon — NoiseTraditional2699 · 2026-08-04
- Google reportedly pre-builds compute for AI self-improvement, with RSI eyed for 2027–2028 — kimmonismus · 2026-08-04