Video: When Algorithms Solve The Wrong Problems — Super Mario Bros Edition

I found this video on a brilliant Twitter thread discussing algorithms that inadvertently solved the wrong problems. It shows how computer code sometimes comes up with unexpected solutions to human-coded problems—in this instance, how to play basic Nintendo video games like Super Mario Bros, and even Pac Man and Tetris.

An algorithm is, simply, a set of instructions usually used to solve a problem or complete a task. Some algorithms can be programmed to learn from a set of provided examples (a training data set) and/or from their own attempts at completing the task or solving the problem, like remembering which route has dead-ends on a maze task. This is described as machine learning.

Although errors in machine learning can often have serious impacts, particularly when the error is not understood, sometimes there is humour in this. For example, an algorithm that was designed to detect cancerous tumours instead taught itself to spot rulers (rulers appeared in all of the training images of cancerous tumours).

Often the errors are the result of problematic human input. But sometimes they are because computers come up with solutions that 'work' but are not what we could have ever imagined in our own rational set of real-world rules.

I recommend watching right until the end of the video to see the programme's ingenious solution to Tetris.