What If We Can't Switch AI Off?

Whenever I picture AI going rogue, I picture the same comforting ending: we pull the plug. These are two thoughts on why that plug might not be there. The first is about where AI will live. The second is about whether it will ever stop.

Prefer to watch? Both parts are short videos. Part 1 is below, and Part 2 is here.

Part 1: What if we couldn't unplug it?

Google is launching TPUs into space to test out data centres in orbit. It sounds like a fascinating idea. We won't have to give up land for data centres, they won't drink up our water for cooling, and they'll leave our energy alone and get theirs straight from the sun. Good, right? Definitely.

I used to picture a future apocalypse (a likely one) where AI goes rogue, and think about how easy it would be to just shut down the data centres. No more GPUs, no more AI. After reading Google's vision and a few other people's takes, I pondered it. What if we couldn't turn off the switch?

Let's say 10 years from now, the most powerful AI lives in data centres in space. If it goes rogue, that's catastrophic. Sure, we'd have missiles and nukes to take them out, but imagine an AI capable of hacking into those very systems to prevent its own extinction.

Cherry on top: AI in physical form is robotics, and companies are racing towards that too. So here's the combination: rogue AI in space + robot army on land. Five years ago, this would've been pure science fiction. Today, it seems inevitable.

Part 2: What if it never stopped?

AI going rogue doesn't sit right with me. It started with prompts that died after answering, then agents, then agents doing long-horizon tasks that run for days or maybe more. But that's about it. A goal that, once fulfilled, makes the AI stop until it's resurrected, or "triggered", by new input. Sure, we can make the goal more time-consuming and stretch its lifespan, but at some point even a rogue AI stops once the goal is achieved.

So I got thinking about a concept: the heartbeat, or pulse, of AI. Give the AI a general goal, say, not dying, and feed it a pulse every n seconds. That pulse could be text or an image, and it acts as the input that keeps it alive. Robots are similar: every frame they see could be that input. The key is that the AI learns when not to reply. If we're going technical, maybe a <silence> token.

That shift is where AI stops waiting to be asked and starts simply existing. Maybe that's what we'll end up calling "conscious", or "AGI".

Parts of this already exist

The pieces are closer than they sound. Agent tools like OpenClaw already have a heartbeat: they wake an agent on a timer, and if nothing needs attention it replies with a "no reply" that gets dropped. Full-duplex speech models like Moshi listen to a continuous stream and stay silent until they have something to say.

What none of them have is the part I keep coming back to: a goal of staying alive, and a pulse that turns a tool you call into something that simply keeps running. Put that somewhere we can't reach, and Part 1 stops being science fiction.