Quick question: Do you want AI to be so well trained, it could help husbands (or wives, for that matter) plan the perfect murder of their spouses? Just as a gut reaction, that feels like a no. I wouldn’t even think it was a particularly hard question.
But America contains many diverse perspectives, and one such perspective was shared by Comma AI founder and longtime jailbreaker George Hotz over the weekend.
The post comes in response to a bunch of big-picture AI alignment plans, most recently the AI 2040: Plan A policy paper from the AI Futures Project. That paper envisions a world in which the world’s researchers collectively choose to slow down AI development for 14 years for the good of humanity. But of course, not everyone who read the paper agrees with its premises or conclusion.
Hotz is in the camp who disagrees. In his post, he argues that the fast-takeoff scenario — the hypothetical where AI rapidly obtains superhuman abilities — doesn’t make a lot of sense. I agree with a lot of what he says here! For Hotz, the best approach to AI alignment and safety is to focus on locally controlled AI models that are closely aligned with the interests of their users.
That’s a cool idea, particularly since it reminds me how much of current AI is built around centrally managed services like Claude and ChatGPT. There are infrastructure-related reasons why AI services developed this way: It’s expensive to host these large state-of-the-art models and most people aren’t using them enough throughout the day to justify truly personal AI. But those factors become less important as the technology develops. Part of what was so exciting about OpenClaw was this experimental, DIY approac …