Aviation-grade rules, open models, and armed robots

This issue makes the case for regulating AI with the rigor aviation demands, examines why the open weight debate is not a simple choice between countries or licenses, and revisits the arguments over autonomous weapons.

A Bloomberg opinion piece proposes regulating AI the way the Federal Aviation Administration regulates aircraft, and Joe agrees with the premise while adding that cybersecurity and self-improving systems worry him as much as bioweapons, and that AI chips are easier to control than downloadable models. (Joe)
Aviation reaches roughly one accident per twenty million flight hours after decades of engineering discipline, while AI companies nearing trillion-dollar valuations still operate closer to startup culture and have adopted nothing comparable. (Joe)
Open weight models, which can be downloaded but whose training data and code may stay secret, are often confused with fully open source ones, and recent Chinese releases have drawn claims of leadership that come without published evidence. (Joe)
After an autonomous attack on Hugging Face, the company hit refusal safeguards on leading American models and ran a Chinese open weight model on its own servers instead, a case that complicates arguments on both sides of the restriction debate. (Joe)
A company building weaponized robots already holds a Pentagon contract, and Alana notes that keeping a human in the loop may mean little in practice, while military secrecy would remove the oversight needed to catch deceptive model behavior. (Alana)