A new open-source benchmark called DishwasherBench wants to know if today's top AI models can figure out how to load a dishwasher.
The project showed up on GitHub and made the rounds on Hacker News, where it picked up 45 points and 19 comments. Rather than testing code generation or trivia recall, DishwasherBench poses a spatial planning problem: deciding where plates, bowls, and utensils should go, and in what order, so everything fits and actually gets clean. It comes from Composo, a startup that builds evaluation tools for AI models. The task sounds almost too ordinary to count as a benchmark, which is exactly why it works as one.
Chatbots have gotten remarkably good at writing essays and debugging code, but the physical common sense a six-year-old has for stacking dishes is a different skill entirely. A benchmark built around a household chore is a sharper probe of that gap than yet another coding or math leaderboard, because there's no shortcut for knowing whether a bowl fits under the rack.
Expect more of these oddball, everyday-task benchmarks as researchers keep testing whether AI models are actually useful around the house, not just impressive in a chat window.