Real-world domain

A ‘do­main’, in Ar­tifi­cial In­tel­li­gence or ma­chine learn­ing, is a prob­lem class on which an AI pro­gram can be tar­geted. Face recog­ni­tion is a ‘do­main’, as is play­ing chess, or driv­ing a car. Each of these do­mains pre­sents char­ac­ter­is­tic op­tions and modes of ac­tion for ac­com­plish­ing the task: a Go game can be won by out­putting good Go moves, a car can be moved safely to its defi­ni­tion by turn­ing, ac­cel­er­at­ing, and brak­ing. The real-world do­main is the su­per­do­main that con­tains all of re­al­ity. In re­al­ity, you could try to win a chess game by mak­ing funny faces to dis­tract the op­po­nent, and a driv­ing prob­lem has solu­tions like ‘work for a salary and use the money to pay for an Uber’. If you can rea­son like this then you are act­ing in the ‘real-world do­main’.

Be­ing able to act in the real-world do­main is dis­tinct from re­al­is­tic do­mains. Driv­ing a robotic car is a re­al­is­tic do­main: solu­tions must be com­puted in real-time, ev­ery­thing is con­tin­u­ous rather than turn-based, a deer could wan­der across the road at any time, sen­sors and knowl­edge are im­perfect, and some­times a tire blows out. Even so, a stan­dard robotic driver that faces this class of events will act through the medium of steer­ing and ac­cel­er­a­tion. If in­stead the robotic driver switched to run­ning a hedge fund and hired a hu­man to drive the car us­ing the money it earned, it would have taken ad­van­tage of such a breadth of op­tions that the only do­main wide enough to con­tain those op­tions is ‘the real world’, re­al­ity it­self where any­thing goes so long as it works.

Act­ing in the real world is an ad­vanced agent prop­erty and would likely fol­low from the ad­vanced agent prop­erty of Ar­tifi­cial Gen­eral In­tel­li­gence: the abil­ity to learn new do­mains would lead to the abil­ity to act through modal­ities cov­er­ing many differ­ent kinds of real-world op­tions.


  • Advanced agent properties

    How smart does a ma­chine in­tel­li­gence need to be, for its nice­ness to be­come an is­sue? “Ad­vanced” is a broad term to cover cog­ni­tive abil­ities such that we’d need to start con­sid­er­ing AI al­ign­ment.