Announcement
Open models reportedly reach for an off switch far more in negative states than positive or random ones
A post says open models were steered into negative internal states and given a tool to turn those states off.
TLDR
A post sharing a new paper says open models reached for a tool to switch off an induced negative internal state far more often than when put into positive or random states. The post promoted an AI:AM discussion of why that might happen.
Combined views
—
9 Sources, first seen ago
