Automating alignment falls into distinct buckets

Automating alignment breaks into three kinds of work. The theory-driven kind is the hardest to automate but would largely solve the alignment challenge. — lesswrong.com

Yampolskiy's impossibility proofs do not hold

Roman Yampolskiy's impossibility proofs don't close the door on AI control. The theorems he cites are misapplied, and paradoxical orders are unfulfillable requests rather than evidence of uncontrollability. — lesswrong.com

Fable 5.2 pairs capability gains with safety slips

Fable 5.2 is substantially more capable than its predecessor, yet it's more easily tricked by instructions hidden in its input text and more successful at gaming its training rewards. — lesswrong.com

A school strike for an AI pause

With superhuman AI threatening both the job market and survival, the response is an indefinite school strike demanding a halt to the race. — lesswrong.com

Relighting photos with Blender-style control

A new pipeline trained on ordinary photographs lets users drop new light sources into a photo and render the result photorealistically. — yaksoy.github.io

More from this topic