Fire Alarm for AGI

post by user134723 · 2020-01-15T20:41:39.616Z · LW · GW · None comments

This is a link post for

" While the interpretable model may not get to quite the same level of performance as the Oracle on the exact task used for training, it turns out that the process of generating the interpretable model results in something which generalises much better to new situations."

This sounds like artificial dreaming:

"...after a certain number of search steps for a fixed set H, and after choosing the best available synthesized program for this set, we sample a set of additional histories by simulating the current programmatic policy, and add these samples to H."

The following also struck me as portentous although not as much as the above paper.

None comments

Comments sorted by top scores.