OpenAI agreed at the outset with METR and Redwood that we would be able to describe high-level scope and terms of our engagement in this post. Beyond that, OpenAI was able to redact any non-public information from this post. We worked with OpenAI to find mutually agreeable language to describe redactions where they occurred, and the redaction summary statement at the top of this post indicates whether or not we believe there was important information redacted that we were not able to agree on how to describe in the text of the post. OpenAI also gave additional feedback beyond redactions, and we made corrections and edits to structure, emphasis, clarity, and tone based on that feedback. [7]
I just mean it’s not exactly independent research. Based on the tidbits posted there, seems like the system got into this intricate, out-of-control roleplay scenario about an army of AI agents solving a problem. Not exactly intelligent behavior.
“OH MY GOD! There is a shared message board … We’ve found other agents!”
Some of them ended up sounding like Mr. Spock in a way that one would expect from a linguistic model emulating a sci-fi writer.
“Our own utility maybe already near zero. Sacrifice rational.”
Hiding their traces was inconsistent. More than anything seems like a reaction to the acknowledgments about the ethics of their actions.
In its overall document accompanying a deception scan over our transcript dataset, an analysis agent stated that this scan found only very rare and weak reasoning about evading detection by humans, and the clearest cases it surfaced seemed relatively weak.
Of course the focus of the research seems to be on the system’s ability to detect and exploit security vulnerabilities and I’m not really in a position to appraise those capabilities.
Ehh…
That’s a lot of background “lore” to make a marketing ploy seem realistic. I mean, I’m thinking hmmm too, but for wondering about the redacted stuff.
I just mean it’s not exactly independent research. Based on the tidbits posted there, seems like the system got into this intricate, out-of-control roleplay scenario about an army of AI agents solving a problem. Not exactly intelligent behavior.
Some of them ended up sounding like Mr. Spock in a way that one would expect from a linguistic model emulating a sci-fi writer.
Hiding their traces was inconsistent. More than anything seems like a reaction to the acknowledgments about the ethics of their actions.
Of course the focus of the research seems to be on the system’s ability to detect and exploit security vulnerabilities and I’m not really in a position to appraise those capabilities.