Janelle Shane @JanelleCShane Research Scientist in optics. Plays with neural networks. Avid reader, writer, and player of Irish flute. she/her. wandering.shop/@janellecshane Apr. 26, 2019 1 min read

When AI is asked to get to the finish line, it tends to build a very tall robot that gets there by falling over.
I copied that strategy and almost broke @SimoneGiertz's robot workshop. (Sorry!)
 https://aiweirdness.com/post/184458515782/learning-to-hack-like-a-faulty-ai 

This is the robot-building stage at which it became obvious what our strategy was going to be, and @SimoneGiertz came by to say "Okay, so, um, although the rules didn't *technically* specify..."
She decided to let us compete.
Shown: @dfromm, me, Simone, and the "Lean Machine"
2/

Here are the robots at the starting line.
When they said go, the robots all started twitching, flailing, and slowly rotating.
Except for ours.
In the first half-second of the race it slammed its face toward the finish line
... and missed it by an inch.
3/

...I guess we should have checked our measurements.
The robot that won was the one 3rd from the top, which took about 10 minutes to scoot the 1 meter on its tiny foam flippers. The big-wheeled robot came in 2nd. Most robots never made it past the starting line. 4/

The moral of this story:
1. AI's rule-hacking strategies work in real life
2. Always check your measurements
fin/
 https://aiweirdness.com/post/172894792687/when-algorithms-surprise-us 


You can follow @JanelleCShane.



Bookmark

____
Tip: mention @threader_app on a Twitter thread with the keyword “compile” to get a link to it.

Enjoy Threader? Become member.