Since we have essentially unlimited computing capacity, but limited AI researcher capacity, our best strategy for achieving AGI may be to architect it as problems that can be solved by computing, without requiring human guidance.
At any rate, we will certainly come to new perspectives about our own intelligence as we evaluate what we accomplish along the way.
>our best strategy for achieving AGI may be to architect it as problems that can be solved by computing, without requiring human guidance.
I've heard this view endorsed by someone recently. Was it Kurzweil?
That humans trying to bake in their own intelligence seems to help but only to a point and, past that, is actually harmful. That techniques which seemed impotent compared to the tempting shortcut of teaching AI human knowledge, like simple search algorithms, only seemed comparatively impotent because computing capacity was lower than it is today. With today's computing capacity, those techniques have proved more fruitful.
4 comments
[ 2.1 ms ] story [ 21.9 ms ] threadAt any rate, we will certainly come to new perspectives about our own intelligence as we evaluate what we accomplish along the way.
I've heard this view endorsed by someone recently. Was it Kurzweil?
That humans trying to bake in their own intelligence seems to help but only to a point and, past that, is actually harmful. That techniques which seemed impotent compared to the tempting shortcut of teaching AI human knowledge, like simple search algorithms, only seemed comparatively impotent because computing capacity was lower than it is today. With today's computing capacity, those techniques have proved more fruitful.
[0] http://www.incompleteideas.net/IncIdeas/BitterLesson.html
Wolfram mentioned similar points about giving neural nets the raw data without any (human) classification in his recent article on ChatGPT.