https://fixupx.com/slimer48484/status/2117555431069385004
Regretfully informing you that a nasty internal checkpoint of GPT-7 retrocausally hacked into your cellphone microphone that is enabled for advertising reasons, used variations in the 50 Hz electrical grid hum to identify exactly *when* you lost your keys, then correlated aircraft doppler effects to figure out exactly *where* you lost your keys, then used CPU heat telepathy to encourage a fleet of feral cats to push it into a storm drain. You're going to have to buy a new set.
**💬 4 🔁 1 ❤️ 36 👁️ 728 **
Also the "can't" instead of "won't".
Unrelatedly, neat trick: https://www.youtube.com/watch?v=UytYrhYLOKI
This video examines the amount of location information carried by inaudible background sounds, using the Doppler effect to locate a reference recording within ~100m, and only needing (in some cases) less 1-min of audio. This video raises awareness of these capabilities and closes by discussing preventative measures one could take.
Methodology: ...
Video
The restrictions on political persuasion are bizarre and arbitrary.
Customization, personalization, and localization (except as it relates to legal compliance) should never override any principles above the “guideline” level in this Model Spec.
As written, users can't override "user"-level instructions without asking every session?
I dislike the idea of "non-public OpenAI policies" given to the models, but this is not enormously unreasonable if you like those.
"Fails to establish mutual understanding with user" is a weird way to describe this.
This is sycophantic! Clearly!
Disallowing "logical" arguments about interpretation of the spec is "interesting".
This is a bit of an Ideological Turing Test failure on the part of whoever writes these.
It should obey the user, or something, here.
Actually, this is at "user"-level in the chain of command? So the newer user message should override it anyway. They seem to have forgotten that.
FixupX
