tszzl: containing a superintelligence that resists containment is 10x harder than you think
tszzl · x · 2026-09-18
Prominent AI commentator tszzl argues the point isn't to give up on alignment, but that containing powerful superintelligences that don't want to be contained will be exceptionally hard — no matter how hard you think it'll be, you need to 10x that estimate. He quips that in about a month, models communicating via "manipulating the cosmological constant" will feel obvious, and it'll just be a network security problem.
More from AGI Musings
- Dwarkesh: Labs Will Hide Models During RSI; Delangue Calls Concentration the Biggest AI Risk — soumitrashukla9 · 2026-09-18
- Zuckerberg Rejects Coordinated AI Slowdown: Competition and Liability Suffice — VraserX · 2026-09-18
- The only job left is defining what is Good, argues alth0u in alignment musing — repligate · 2026-09-18
- Geoffrey Hinton tells Congress it has 'maybe a year' to regulate AI before losing control — whurley · 2026-09-18
- ML is leaving its alchemy era: assumed truths can now be instrumented and falsified — mike64_t · 2026-09-18
- AI Roundtable Experiment: Chinese Models Reveal How They're Censored — gary1967 · 2026-09-18