Short answers, pulled from the story.
Commonsense knowledge in AI refers to everyday facts about the world that all humans are expected to know, such as "lemons are sour" or "cows say moo." It underpins commonsense reasoning, allowing AI systems to make default assumptions and inferences about ordinary situations. Teaching machines this kind of knowledge remains an unsolved problem in artificial general intelligence.
John McCarthy created the first AI program to directly address commonsense knowledge, called Advice Taker, in 1959. McCarthy recognized that reasoning about the world required background knowledge that humans absorb without thinking, not just logic and calculation.
The Winograd Schema Challenge is a modern benchmark test for commonsense reasoning in AI. All existing computer programs that attempt human-level AI perform extremely poorly on it, illustrating how far machines remain from human-level common sense.
ConceptNet is a commonsense knowledge base and natural language processing engine that grew out of the crowdsourced OpenMind Commonsense project. As of 2012 it contained 21 language-independent relations covering connections such as IsA, UsedFor, HasA, CapableOf, and Causes.
BullySpace is an extension of ConceptNet developed by MIT researchers around 2013 to detect taunting comments on social media. It included over 200 semantic assertions based on stereotypes, enabling it to infer whether a comment like "Put on a wig and lipstick and be who you really are" was an insult depending on whether it was directed at a boy or a girl.
At Lawrence Livermore National Laboratory, commonsense knowledge was applied in an intelligent software agent designed to detect violations of a comprehensive nuclear test ban treaty. This represented one of the more specialized real-world deployments of commonsense AI technology.