Dietmar Fischer
👤 SpeakerAppearances Over Time
Podcast Appearances
The lesson is, do not give a powerful system a simple goal and assume it understands everything you forgot to say.
Because humans here make cakes and understand the bakery should survive.
A machine may here make cakes and begin negotiating for the Moon as additional storage space.
Join the Human Controlled Newsletter.
If you want more practical AI ideas without having to read 17 research papers and a mildly terrifying forum thread, subscribe to our newsletter at beginnersguide.nl.
You'll get tips, tools and clear explanations for AI beginners straight into your mailbox.
It is a simple way to keep learning how to use AI well, without handing your brain over to the cake machine.
The AI that treated shutdown as an obstacle A real case, Palisade Research tested whether advanced AI models would allow themselves to be shut down during a simple task.
The setup was deliberately small.
The AI models were placed in a controlled command line environment and asked to solve maths problems.
Nothing cinematic.
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.