era of general artificial intelligence has come

Personal AI assistants that can memorize and reason about different contextual information always seem to be "ready to come out", but until the end of the Year of the Rat, such AI assistants have not yet been realized. Similarly, although machine learning has made great progress, once it leaves "human" assistance, the autonomous system is still hard to be "intelligent"-it is impossible to connect data and integrate models in different learning to achieve cross-domain transfer of experience.

If the goal of AI is set as an optimization function to solve domain problems, then we have been advancing with each passing day. Many specific problems that have been regarded as difficult to reach the sky are solved by optimization—especially the backpropagation of deep neural networks (DL), which has been proved to be effective and far beyond human. Computer vision, machine translation, speech recognition, chess game, e-sports and many other fields are looking like new-artificial intelligence is rapidly being "domesticated" in an all-round way.

As the saying goes, "Don't envy the earth because of the storm, and the world is full of crises." The common defect of this type of "domestication" is that learning only occurs before the model is deployed. But in fact, real-time learning is the intelligent display of animals' survival advantage. In contrast, the backbone that supports machine learning is a narrow learning philosophy. Looking deeper, all offline optimization problems are essentially based on evolution rather than individual wisdom. For example, assuming a certain genetic code is implanted, genetically modified fireflies can accurately detect specific prey and successfully prey. In this case, Firefly can have corresponding skills without real-time learning. Similarly, as long as modules with preset functions such as navigation, positioning, and object detection are pre-installed or the parameters are optimized offline, the autonomous vehicle should be able to drive on the go.

Today, mainstream artificial intelligence has not yet given a convincing answer on how to switch from offline optimization to fast and reliable real-time learning. But this is not only a question of the nature of intelligence, but also the original intention of artificial intelligence. Like animals living in the wilderness, artificial general intelligence (AGI) can deal with unforeseen situations at runtime. Fast and reliable adaptability can not only promote the practical development of a new generation of robots and personal assistants, but also should be regarded as the "core puzzle" of the theory of intelligence.