Seems like if we can keep kids safe from online pornography we can also restrict access to ai bots. Would love more than just lip service from Bonta, but I’m sure he knows who butters the bread and doubt there will be more than a stern warning.
The whole Adam Raine story was heartbreaking. The worst part, to me, was reading some of his writings and the feedback from the chat engine.
There's a known, repeatable failure mode in these engines that anyone who's worked with them for more than a couple hours can tell you: when a conversation goes on too long, it "feeds back" on itself. I don't fully understand the mechanism (I believe it's partly to do with the model's attention mechanism getting saturated and newer content dominating over older content, and it's related to the "jailbreak" solution where you hit the machine with so much text that it "forgets" the directives it's been given by its creators when it booted up). But the end result is that over time, the conversation centers on the more recent topics, ideas, and tones over the initial configuration. That's one of the reasons you can "trick" these models into being racist by talking like a racist, and so on.
Reading the excerpts I've read from the chat history Raine had with his session, it seems pretty clear that it's gone on so long that the session is "reflecting" his own writing and mood back at him. And that's the heartbreaking part: it's, in essence, coaching him to end his life because he's been talking about ending his life for so long in the conversation that it's saturated the model.
It's this poor young man's own pain reflected back on him.
6 comments
[ 0.18 ms ] story [ 19.6 ms ] threadThe "think of the children" crowd should not be given a single inch. Nothing good ever came from it, and by now, I believe that nothing ever will.
There's a known, repeatable failure mode in these engines that anyone who's worked with them for more than a couple hours can tell you: when a conversation goes on too long, it "feeds back" on itself. I don't fully understand the mechanism (I believe it's partly to do with the model's attention mechanism getting saturated and newer content dominating over older content, and it's related to the "jailbreak" solution where you hit the machine with so much text that it "forgets" the directives it's been given by its creators when it booted up). But the end result is that over time, the conversation centers on the more recent topics, ideas, and tones over the initial configuration. That's one of the reasons you can "trick" these models into being racist by talking like a racist, and so on.
Reading the excerpts I've read from the chat history Raine had with his session, it seems pretty clear that it's gone on so long that the session is "reflecting" his own writing and mood back at him. And that's the heartbreaking part: it's, in essence, coaching him to end his life because he's been talking about ending his life for so long in the conversation that it's saturated the model.
It's this poor young man's own pain reflected back on him.