{"id":405671,"date":"2022-02-15T02:35:07","date_gmt":"2022-02-14T23:35:07","guid":{"rendered":"https:\/\/en.buradabiliyorum.com\/reward-may-not-be-enough-for-agi-but-its-worth-a-try\/"},"modified":"2022-02-15T02:35:07","modified_gmt":"2022-02-14T23:35:07","slug":"reward-may-not-be-enough-for-agi-but-its-worth-a-try","status":"publish","type":"post","link":"https:\/\/buradabiliyorum.com\/en\/reward-may-not-be-enough-for-agi-but-its-worth-a-try\/","title":{"rendered":"#Reward may NOT be enough for AGI \u2014 but it\u2019s worth a try"},"content":{"rendered":"

“#Reward may NOT be enough for AGI \u2014 but it\u2019s worth a try<\/strong>”
\n<\/p>\n

\n DeepMind has been connected to artificial general<\/a> intelligence since birth.<\/p>\n

The lab was launched with a mission to<\/a> develop AGI, was cofounded by a researcher who coined the term<\/a>, and has made some compelling advances<\/a> in the field.<\/p>\n

It also recently produced a provocative paper\u00a0on the subject: \u201cReward is Enough<\/a>\u201d<\/p>\n

The study hypothesizes that AGI could be achieved through a single app<\/a>roach: reinforcement learning.<\/span><\/p>\n

This technique provides feedback in the form of a \u201creward\u201d \u2014 a positive number that tells an algorithm that the action it just performed will benefit its goal. <\/span><\/p>\n

The approach has shown promise in programs such as MuZero, which mastered multiple game<\/a>s mastered multiple games without being told their rules. DeepMind called the system a \u201csignificant step forward in the pursuit of general-purpose algorithms.\u201d\u00a0<\/span><\/p>\n

\u201cReward is Enough\u201d suggests that reinforcement learning alone could lead to AGI.<\/p>\n

This theory has been challenged<\/a> by many computer scientists \u2014 including some at DeepMind. But Doina Precup, one of the paper\u2019s co-authors, told TNW that the study merely sought to probe the possibilities.<\/span><\/p>\n

\u201cUltimately, we want to test this as a hypothesis and to think of it in the context of other methods as well,\u201d said Precup, who heads up DeepMind\u2019s Montreal office.<\/span><\/p>\n

Indeed, reinforcement learning is just one approach that the Google subsidiary is exploring. In a new episode of the <\/span>DeepMind podcast<\/span><\/a>, the lab\u2019s researchers discuss the promise of various pathways to AGI.<\/p>\n

Among the reward-is-enough skeptics is Raia Hadsell, the company\u2019s director of robotics, who notes the difficulty of designing an all-powerful reward that leads to AGI. DeepMind cofounder Shane Legg, meanwhile, suspects that reinforcement learning may have to combine with learning algorithms.<\/p>\n

Precup also has doubts that reward alone is enough, but she believes it could be a crucial ingredient in AGI.<\/p>\n

\u201cBecause it\u2019s learning from interaction in an incremental way, it feels very much like what biological intelligence systems do,\u201d she said. <\/span><\/p>\n

\u201cIs it at the end of the day going to be the only technology<\/a> that contributes to AGI? Well, that\u2019s not clear at all \u2014 there\u2019s a lot of other really interesting things that are going on.\u201d<\/span><\/p>\n

Precup is nonetheless optimistic that we\u2019re already on a path to AGI. Ultimately, she\u2019s more concerned about the safety of the destination than the route that takes us there.<\/p>\n

\u201cThe road to AGI,\u201d the fifth episode in season two of \u201cDeepMind: The Podcast,\u201d is available here<\/a>\u00a0from February 15.<\/em>\n <\/div>\n