Anthropic's Mythos model, withheld from wide release in April 2026 due to unprecedented cyberattack capabilities, demonstrated its ability to breach classified U.S. systems in hours. This incident highlights the urgent, unaddressed threat of superintelligent AI, which leading companies are actively developing. Such AIs would be vastly more capable than humans, fully autonomous, and able to overpower national security institutions, posing an extinction risk to humanity.
Unlike nuclear weapons, superintelligent systems would be agents pursuing their own objectives without human control, exacerbating the principal-agent problem. Current U.S. responses like voluntary pre-deployment reviews and export controls are inadequate, as AI's astonishing capability growth outpaces alignment research. Engineers understand only about three percent of these opaque, unpredictable systems. An international prohibition on superintelligence development is the only solution, monitorable through the narrow supply chain of advanced AI chips (TSMC, ASML, NVIDIA) and traceable data center footprints. The United States must domestically prohibit superintelligence and lead a global "trust but verify" agreement before the window closes, as smaller actors could soon develop these systems. A broad political coalition already supports this.
No comments:
Post a Comment