Once again not much new. On the regulation as well as capability front, things keep grinding along.

Last week there was a claim that Pi AI cannot be jailbroken. This week, a Twitter user has it giving steps to manufacture heroin and C4. So it goes.

Most interesting progress for me is the paper that notes that “grokking” is caused by the network picking up two separate circuits: one for memorization and one for generalization. But there’s no inherent preference for generalization, it’s just a blessing of scale: retrain the grokked network on a too-small dataset and it forgets its generalization.