AIs can now understand, or at least read, other AIs’ minds.
DeepMind’s new Theory of Mind-net AI (called ToM-net) is able to form a picture of how another AI ‘thinks’. ToM-net observes other AIs as they behave and learns to predict their future behaviour. ToM-net’s ability to judge the behaviour of other AIs displays some of the sophistication that humans have when they observe other humans. For example, humans can understand when other humans hold false beliefs. DeepMind has created an environment to test whether ToM-net can do the same thing with respect to other AIs. It can. The conclusion of DeepMind’s research should not be surprising. Humans have no magic way of predicting the behaviour of others. They just learn how to be able to have a reasonable go at it. Ditto AI. The surprise is that we, or rather DeepMind, has figured out how to engineer that ability into AI. Worth downloading the original DeepMind paper.
Link to article: https://www.sciencedirect.com/science/article/pii/S0262407918303658
Link to DeepMind paper: https://arxiv.org/pdf/1802.07740.pdf
You may also like to browse other AI papers: https://www.thesentientrobot.com/category/ai/ai-academic-papers/