One-minute frequency disturbances from unintended synchronization in cyber–physical power systems
Some cool people stared at oscillations of our power grid frequency, spotted a notable repeating pattern every minute and went digging into it.
Some cool people stared at oscillations of our power grid frequency, spotted a notable repeating pattern every minute and went digging into it.
Instead of patching individual CPU instructions, people now patch individual LLM neurons for behavioral changes.
The post framing is nonsense (nothing to do with abliteration), but someone actually put in the work and proved that these backdoors in weights are not just possible but easy and cheap even.
The same stuff could be in the base model, but, it could also be in the inference engine. Or in the firmware of the GPU a model runs on. Or in the silicon.