On: LoRA Fine-Tuning Efficiently Undoes Safety Trainin...

[cantSpellSober]

I, too, clicked through thinking this was about LoRaWAN (despite context clues)

[tlavoie → cantSpellSober]

Same, and I was hoping to be highly amused by someone's Arduino somehow unlocking heavily computational LLM secrets via tiny tiny dribbles of radio traffic.

[McNutty → tlavoie]

Same also. I was fascinated with the idea that someone was trying to fine-tune some RF parameters using AI and accidentally undid some safety mechanism

[sleepybrett → cantSpellSober]

The context clues being that every VC in the world is now trying to prop up their over commitments in 'AI' ...