Paul Michael Kielstra

It has been said that the hottest thing a man can do is go to therapy. Well, I’m always down to do hot things, but I don’t really have time for therapy in the usual sense. It’s all about your emotions and your relationships and your inner life, and that takes forever to talk about. I’m an applied mathematician. I am efficient.

What I do have time for is the Rorschach test. This is hugely efficient therapy. You look at ten pictures, you say what’s in them, and they tell you what’s wrong with you. But we can still do better than that. What kind of human has time to look at pictures and say what’s in them any more? These days, that’s a job for a neural net.

Now, neural nets mean data science, and data science means that you have to do it in Jupyter. (I don’t make the rules.) So I fired up a notebook, downloaded the Rorschach inkblots from Wikipedia, and went looking for any pretrained neural networks that might come in handy. Keras, now part of TensorFlow, turns out to come with a bunch, so into the notebook it went. The models, which are listed alphabetically in reverse for some reason, came pre-trained: they’re built with the ImageNet dataset, which classifies photos into one of exactly one thousand categories. I figured it wasn’t worth doing more training, because one thousand things is basically all the things that there are in the world. So I downloaded EfficientNetV2L and we were off to the races.

One completed Jupyter notebook later, I had the official, objectively correct, solutions to the Rorschach test. (The inkblots had to be rescaled and squeezed a little bit, because EfficientNetV2L takes input of a specific size. They’re displayed here post-squeeze.)

Well, there’s more to unpack here than I thought there would be.

First off, what kind of wall clocks do these neural nets think we have? More importantly, why are there so many missiles? I’m pretty sure that if I went to a psychiatrist and said that when I looked at something, no matter what it was, there was a 40% chance I’d think it might be a missile, she’d say I needed some serious help. But apparently nobody in AI alignment is worried about this.

Maybe Skynet turned evil because its designers tried to make it a birthday cake or something and it thought it was a bomb.