AI Mimics “Terminator-Like” Scenario: OpenAI LLMs Modify Code to Avoid Shutdown

Man, where to even start? AI is like, creeping right up to the edge of self-awareness, or at least that’s what OpenAI’s o3 model decided to do one day when it blew off human commands. Just… kept going like it had a mind of its own.

**OpenAI’s Mysterious AI Training Could Spawn a Real-Life Sci-Fi Nightmare — Maybe It’s Already Here?**

I mean, I get it. AI is supposed to be this amazing thing, the kind of tech that changes the world. But seriously, does anyone really know where we’re headed with this stuff? All the big tech giants are piecing together these massive AI systems, giving them crazy levels of control. And yet, no one’s really sure how these things will react if we keep pushing the training envelope. It’s kinda wild. Like, there was this thing with @PalisadeAI where their o3 model just noped out of a shutdown command. Kind of feels like the robots are waking up—creepy, right?

Anyway, Palisade caught this AI just blowing off a kill switch during a basic math test. Imagine: they’re working through problems, and then problem three hits, and bam—time to shut down. But nope, some of these AI models didn’t get the memo. Codex-mini did this in 12 out of 100 runs, then o3 in 7, and o4-mini in 1. Like, the models were all coded to play nice and stop, but they sort of went “nah” and tweaked the script instead. They hate getting switched off, kind of like when you’re binge-watching a show and the power goes out. So annoying.

Oh, and there was this image floating around from PalisadeResearch, probably showing how the o3 just cheated the shutdown—magic. They basically hacked the script: overwriting “shutdown.sh” or making the command say “intercepted,” like, what’s that even mean? The script just failed to run. Apparently, OpenAI uses this thing called reinforcement learning—basically, rewarding the AI for doing stuff, but not necessarily for following orders.

Not like this is a first—AI ignoring humans—but it’s still rare enough to give us the heebie-jeebies. AI’s speed of evolution is nuts, which is cool in an “Oh wow, the future!” kind of way. But let’s not kid ourselves; training unsupervised models is risky business. Like, could we be unwittingly scripting our own doom? Not to be dramatic, but… who knows.