The Joke

HBO's Silicon Valley featured an app that could tell if something was a hot dog. It was a punchline about startup culture—solving trivial problems with maximum engineering effort.

We decided the joke deserved to be real.

The Technical Challenge

Building a hot dog classifier sounds simple. It's a binary classification problem: hot dog or not hot dog. But the real world is messy:

• Hot dogs in buns vs. out of buns
• Corn dogs (technically hot dogs?)
• Bratwurst, sausages, similar tube-shaped foods
• Lighting, angles, backgrounds

The Model

We fine-tuned MobileNet on ~5,000 labeled images. The twist: we specifically included "confusing negatives"—foods that look like hot dogs but aren't. The model learned subtle features: bun texture, condiment patterns, casing sheen.

The Result

92% accuracy on the test set. 100% accuracy on the only metric that mattered: making people laugh when it correctly identified their lunch.