After months of AI experts investigating the usefulness of virtual children for poor future parents, scientists have created an AI baby. But what could this artificial intelligence baby possibly be for?
Did scientists create an AI baby?
Scientists at Princeton University decided to take a new approach for training artificial intelligence. Instead of building from scratch and absorbing advanced datasets, they decided to train the program as if it was a human baby.
The AI baby project was taught general physics, the same way we teach children. For example, the triangle goes in the triangle hole, the square goes in the square hole, et cetera.
To do this, the scientists have the newborn AI visual animations of objects such as a ball bouncing off a wall. Afterwards, the program is tasked with creating its own animations that follow basic physics.
In the study, the AI baby method was able to learn faster and more consistently than traditionally trained AI. For comparison, typical AI models have preloaded datasets before they learn their intended purpose. E.g. GPT-3 knowing how to speak before speaking about a specific topic.
The study discovered that the typically way of training artificial intelligence is considerably less ideal. However, training AI in this new method may only be suited for some forms of AI, not all use cases.
Will this be the basis of virtual children?
The idea of an AI baby has been a headline puller as of late. However, those stories have been more about actual virtual children instead of programs designed to learn like this.
Of course, this method of training may end up being the basis for virtual Metaverse children in the future. After all, there have already been projects such as Replika that learn from our actions. If the AI model is esssntially a blank slate, that could work wonders for virtual children.
The idea of an AI baby for families is still very dystopian. However, figuring out how that technology could come about is still very intriguing.