I’ve been following AI development for 40 years, since discovering ELIZA in high school, and invested in and implemented machine learning for 15. From my vantage point, an uncontrolled or misaligned superintelligence escaping human oversight is a real risk. But I think the classic “AI takeover” story still requires capabilities we just haven’t demonstrated. Sustained autonomous agency, the ability to evade oversight and preserve access, and capability growth that outruns our ability to maintain effective control are some of the precursors to that apocalyptic vision. Those aren't Hollywood props. They also aren't here yet, at least not at an effective level. And that's where serious people split. Some of the researchers closest to frontier models think we're closer than the public debate admits. Others think scaling current methods will plateau short of that. Who’s right? I don’t know. What I do know is that the probability is not zero, and that “not zero” on this class of outcome is already an investment and policy problem. The scenario I treat as the base case is not some conscious mind rebelling. It’s general-purpose models deployed in increasingly autonomous, task-specific systems producing cascading cyber-physical damage. Think force multipliers for human actors, or agents whose failures propagate through interconnected infrastructure. That mess can look like collapse without anyone building a god. If a true loss-of-control event happens, I expect it to arrive through that channel first: agents, tools, networks, access and incentives. Not Skynet. And, in my second career as a screenwriter, I know a little about movie villains, too.
— Will Bermender (@WillBermender) Sep 14, 2026
from Will Bermender on Twitter
September 14, 2026 at 09:41AM
