It’s Elon Musk’s worst nightmare: a group of computer researchers are building a self-aware AI, programming their creation with an insatiable desire for material wealth and teaching it how to kill. Thankfully, the AI’s name is Mario; he’s looking for coins, and his enemies are Goombas. This is the "Mario Lives!" project, a quirky bit of research from the University of Tübingen in Germany that aims to build a "Living and Conversing Mario Agent."
Created as an entry for an annual video competition organized by the Association for the Advancement of Artificial Intelligence (AAAI), "Mario Lives!" allows players to program the famous mustachioed plumber with information and motivations using natural speech voice commands — instead of having to fiddle with a game controller.
Give Mario knowledge of life and death by telling him: "Goomba dies when you jump on goomba."
"As most of you know, this is Mario," explains a researcher in the video. "But what you do not know is that this Mario has become aware of himself and his environment — at least to a certain extent." A second researcher then shows how Mario can be taught pieces of information (e.g. "Goomba dies when you jump on Goomba") and given motivations that shape his behavior (when Mario is "hungry" he looks for coins; when he’s "curious" he explores his world). Mario responds by repeating back any new knowledge in a computer-generated voice — an effect that's a little eerie when he's told "don’t be so happy" and replies "Somehow, I feel less happy."
Fabian Schrodt, one of the researchers involved in the project, tells The Verge that as far as artificial intelligence goes, "Mario Lives!" is comparable to a regular video game opponent. However, he says, it's the combination of AI programming that learns and adapts as well as the addition of principles from psychology that's interesting. Schrodt adds that the same team is also working on a follow-up project where both Mario and Luigi are AI-controlled and are able to speak to each another using computer-synthesized voices. That way, says Schrodt, they’ll be able to gather new information independently and then share it between them, teaching one another as they go. Maybe Musk should be worried after all.
Verge Video Archives: Can We Build a Conscious Computer? (The Big Future, Ep. 9)