I picked up Janelle Shane’s book, You Look Like a Thing and I Love You, a couple of weeks ago. Combining clear explanations of AI with hilarious examples of how weird it can be, this was one of the best books I’ve read in a while.
What is this book?
You Look Like a Thing and I Love You is a book about “AI Weirdness,” inspired by Dr. Shane’s blog of the same name. In it, she explains how AI works in clear, accessible language; and she also gives us dozens of examples of how it fails, often in spectacularly funny ways.
The funny bits
The title, like many other examples in the book, was created by a neural net that Dr. Shane fine-tuned. This one was trained on a set of pickup lines, but throughout the book she gives examples of NN-generated ice cream flavors, desserts, D&D spells, and metal bands.
I loved this book primarily because it made me laugh so much. The AI generated outputs (and the author’s skill in handpicking the best ones) are frequently hilarious. Consider these ice cream flavors:
- Chocolate Peanut Chocolate Chocolate Chocolate
- Beet Bourbon
- Beast Cream
- Praline Cheddar Swirl
Another of my favorite examples was from “intermediate transfer learning”—where she fine-tuned a net on ice cream flavors, then again on metal bands. The results in the middle were a mix of the two:
- Dirge of Fudge
- Death Cheese
- Blood Pecan
- Silence of Coconut
- Spider and Sorbeast
- Blackberry Burn
There were humorous stories, too. Reinforcement learning agents working within simulations would exploit the imperfect physics to clip through the ground or eat children that they could produce at no cost. An evolutionary robot trying to direct traffic to one of two hallways learned to evolve into a wall.
All throughout the book, Dr. Shane explained some of the weirdest bits of AI in hilarious ways. The robot illustrations were adorable, too!
Accessible explanations
I also loved how accessible the explanations were. The book started with generic computer programs and built up to how AI works, then explained (through a great example about “learning what makes a sandwich tasty”) how neural nets learn through gradient descent.
The book also covered reinforcement learning, decision trees, transfer learning, AI bias, overfitting, adversarial attacks, and far more. And they were all so well-written!
Closing thoughts
I loved this book. It clearly and accessibly explained how AI works, and how it doesn’t, with hilarious examples and cute illustrations. I recommend this to anyone in the data-science-adjacent world.