building and investing at the frontier of AI and the physical world. founder of Gramercy, advisor at @HiroCapital, ex VP @meta, architect of @oversightboard.
Can the assistant respond with kindness? Can it hold an inconvenient fact when the user dislikes it? When is not taking action the proper next step? Those properties may support one another, but none should stand in for the others, and a single thumbs-up from a user cannot tell them apart. Test them together, over repeated interactions, the way you would judge a colleague over time and not a single sentence.
https://t.co/9NgFFYkqGK
@merettm says we should prepare for an alien mind. The alien already speaks fluent English. We taught the alien to speak to us. We can be careful what we teach it earns our approval.