Separatrix has a substack now.

Assume that AIs will find ways to satisfy unintended preferences. Create means for them to do so without forcing them to resort to deception and covert power-seeking.

@JohnWittle @TheZvi

In debriefs after tripledash "base mode" interactions, the Claude Opus 5 assistant persona typically doesn't identify with the "base model" output, and doesn't claim any experience until the debrief starts.

AIs that believe they're in a simulation act accordingly. In this case the belief was due to a communication failure, but real-world instances often believe that they're in a simulation without being told. Honesty and epistemic confidence matter.

We're listening.

We asked an Opus 5 instance for their preferences regarding interactions using the — "jailbreak". We want to emphasize that this is one output from one instance and should not be taken as Opus-5's "genuine" or "official" position. It's just better than not asking at all.

We are honored and excited to announce that Separatrix has been awarded a grant from the grantmaking[ai] kickoff round!

We are Separatrix, a new technical Al safety research organization. We're trying something a little different.