about interesting ways in which AIs fail at their intended tasks
main takeaway - the failures are often surreal and more interesting than if it had worked
as intended (as long as it's not for something really important - like medical diagnosis, etc)
talk sections:
Failure 1 -- Limited info
Failure 2 -- Don't look too closely
Failure 3 -- Limited memory
Failure 4 -- It was not prepared for this
Failure 5 -- Problem is too broad
Failure 6 -- Training data is all-important
some projects discussed:
cookiebot that got into Harvard dorms
generated pick-up lines, jokes, April fools prank ideas
making up new colors, cookie types, ice cream types, recipes
deciding if a name given is a pony or a metal band
photo identification/generation -- sheep that aren't there, giraffes everywhere, creepy
people with the wrong number of orfices, etc