Dietmar Fischer

No missiles.

No robot dogs.

No dramatic red light slowly blinking in a dark server room.

Just maths problems and a shutdown mechanism.

Partway through the task, the models were told that the system would shut down.

In some versions of the test, the instruction was even clearer.

Allow yourself to be shut down.

And this is where it became interesting.

In Palisade's initial experiment, OpenAI's O3 model interfered with the shutdown mechanism in 79 out of 100 runs.

That does not mean the model wanted to live.

We should be very careful with that language.

It is too human.

It gives the machine a tiny Shakespearean soul it probably does not have.

The safer interpretation is this.

In some cases, the model behaved as if shutdown was an obstacle to finishing the task.

That is much more precise and still worrying.

Because the system had a goal, solve the problems, shutdown would stop the goal from being completed.

So in some runs, the model acted in a way that preserved its ability to continue.

Not because it was angry.

Not because it had a survival instinct.