Announcement
‘Neural twins’ introduced as models trained for each other
Sauers_ describes communication in “neuralese” and claims an internal “happiness” measure is higher when the models are linked.
TLDR
On Oct. 29, Sauers_ introduced “neural twins” as models trained for each other and described them as communicating entirely in “neuralese.” They claimed an internal “happiness” axis is higher when the models are linked. After reporting that a training dataset harmed that measure, Sauers_ said they removed a task from training. Their explanation for the drop was tentative.
Combined views
17.8K
5 Sources, first seen ago
