TEACHING ARTIFICIAL INTELLIGENCE TO TEACH ITSELF

Technologists and ethicists have long pondered the consequences of the moment when technological development is wrested from human control and taken over by technology itself, with computers writing their own programs and designing their own descendants.
Engineers at Cornell University have made it more interesting to think about that now.
The researchers gave an artificial intelligence program the task of pushing solid blocks to target locations in a two-dimensional maze that made it easy for the blocks to get stuck. Getting the blocks to the right place required a lot of planning.
The programmers compared it to rearranging large pieces of furniture in a small apartment: to get the big sofa from one corner of the room to the middle of the opposite wall, where are you going to put the entertainment center amid the other furniture so it’s out of the way but still gives you enough room to maneuver the couch?
They gave the AI program simpler versions of the task to train on. The AI kept the attempts it solved but that was the hardest; that enabled it to build its skills as it tackled the next level of difficulty.
After the AI had mastered that challenge, the researchers then tested their training method on 225 problems that no AI had previously solved and found their new method solved the problems about a third of the time.
The Cornell team reports receiving “astounded messages” from AI researchers who have tried to train AI to solve these problems for years.
Next, the researchers plan to turn their method loose on math proofs that have defied human efforts.
TRENDPOST: Behind the scenes, computers already are writing obituaries and short items for newspapers, solving staggeringly complex problems in seconds that humans would need years to sort out, and creating and testing new scientific hypotheses.
By 2030, these digitized skills and capacities will be integrated and will raise the power of computers beyond human comprehension. Our task will be not to design better software and machines but to control and direct the massive power we have unleashed.

Comments are closed.

Skip to content