What surprised me about You Look Like A Thing And I Love You is that it’s genuinely funny – laugh-out-loud-funny, read-quotes-to-your-family-over-breakfast-funny. Who would not be charmed by an AI that develops My Little Pony names and suggests “Parpy Stink” and “Starsh*tter”? Or the accidental Murderbot that was supposed to be acting as a friendly usher? Or the curiosity-driven AI that plays Pacman by going to watch the ghosts, because they’re so interesting?
It’s not just a bunch of silly-AI gags, though. There may be a My Little Pony called Raspberry Turd on every other page, but there’s also a great deal of information about how machine learning actually works and why it finds certain kinds of problem a lot more difficult than others. Janelle Shane runs through various sources of AI-weirdness: AIs being trained in simulations (because simulators are faster and safer) and then finding ways to hack the simulation; AIs being fed subtly flawed training data (such as the AI which noticed that the difference between cancerous skin and healthy skin is that there’s usually a ruler in the picture when skin cancer is involved); AIs seeing giraffes everywhere in photographs of the savannah (because people like to take photos of giraffes, it’s safest to assume there’s one in the photo).
I learned a lot and laughed a lot.
[jetpack_subscription_form title=”Receive these posts by email” subscribe_text=”(You can unsubscribe at any time)” subscribe_button=“Subscribe” show_subscribers_total=”0”]