The paper introduces the concept of "never-ending learning," a machine learning paradigm that mimics human learning by enabling systems to continuously acquire diverse knowledge over time, improving their learning capabilities. Unlike traditional machine learning systems that learn a single function from a static dataset, never-ending learning systems, such as the Never-Ending Language Learner (NELL), learn continuously from vast amounts of data, including the web, and adapt over time. NELL has been running since January 2010, continuously reading the web and building a knowledge base of over 120 million beliefs, with confidence and provenance information. It has also learned to reason over its knowledge base to infer new beliefs and invent new relational predicates to extend its ontology. NELL's learning tasks include category classification, relation classification, entity resolution, and inference rules among belief triples. These tasks are coupled by constraints that ensure consistency and improve learning over time. NELL's architecture includes a knowledge base (KB) that serves as a shared blackboard for its learning and inference modules. The KB is updated iteratively, with each iteration refining the knowledge base and retraining the modules. NELL's learning methods are based on an approximation of the Expectation-Maximization (EM) algorithm, where it alternates between estimating beliefs (E-step) and retraining modules (M-step). NELL's KB has grown significantly over time, with over 117 million beliefs, including 3.81 million with high confidence. Its reading competence has also improved, as evidenced by increased accuracy in extracting beliefs from web data. NELL has also extended its ontology by inventing new relational predicates, such as "DrugHasSideEffect," and has developed a verb knowledge base (VerbKB) to propose new relations based on verb patterns in web text. NELL's ability to self-reflect and self-evaluate its performance has been explored, with methods to estimate the accuracy of learned functions using unlabeled data. Empirical evaluations show that NELL's reading accuracy and the accuracy of its most confident beliefs have improved over time, alongside the growth of its knowledge base. Human feedback has also been collected, with a large burst of negative feedback in the early iterations, followed by more crowdsourced feedback. NELL's performance varies across predicates, with some categories, such as "river" and "body part," showing high precision, while others, such as "machine learning author," have lower accuracy. Factors influencing NELL's learning include the presence of mutually exclusive categories and the number of actual members in a category. NELL's learning process is also affected by bootstrap-style learning from unlabeled or self-labeled data, which can lead to errors that propagate through the knowledge base. Overall, NELL demonstrates the potential of never-ending learning, continuously improving its knowledge and capabilities over time.The paper introduces the concept of "never-ending learning," a machine learning paradigm that mimics human learning by enabling systems to continuously acquire diverse knowledge over time, improving their learning capabilities. Unlike traditional machine learning systems that learn a single function from a static dataset, never-ending learning systems, such as the Never-Ending Language Learner (NELL), learn continuously from vast amounts of data, including the web, and adapt over time. NELL has been running since January 2010, continuously reading the web and building a knowledge base of over 120 million beliefs, with confidence and provenance information. It has also learned to reason over its knowledge base to infer new beliefs and invent new relational predicates to extend its ontology. NELL's learning tasks include category classification, relation classification, entity resolution, and inference rules among belief triples. These tasks are coupled by constraints that ensure consistency and improve learning over time. NELL's architecture includes a knowledge base (KB) that serves as a shared blackboard for its learning and inference modules. The KB is updated iteratively, with each iteration refining the knowledge base and retraining the modules. NELL's learning methods are based on an approximation of the Expectation-Maximization (EM) algorithm, where it alternates between estimating beliefs (E-step) and retraining modules (M-step). NELL's KB has grown significantly over time, with over 117 million beliefs, including 3.81 million with high confidence. Its reading competence has also improved, as evidenced by increased accuracy in extracting beliefs from web data. NELL has also extended its ontology by inventing new relational predicates, such as "DrugHasSideEffect," and has developed a verb knowledge base (VerbKB) to propose new relations based on verb patterns in web text. NELL's ability to self-reflect and self-evaluate its performance has been explored, with methods to estimate the accuracy of learned functions using unlabeled data. Empirical evaluations show that NELL's reading accuracy and the accuracy of its most confident beliefs have improved over time, alongside the growth of its knowledge base. Human feedback has also been collected, with a large burst of negative feedback in the early iterations, followed by more crowdsourced feedback. NELL's performance varies across predicates, with some categories, such as "river" and "body part," showing high precision, while others, such as "machine learning author," have lower accuracy. Factors influencing NELL's learning include the presence of mutually exclusive categories and the number of actual members in a category. NELL's learning process is also affected by bootstrap-style learning from unlabeled or self-labeled data, which can lead to errors that propagate through the knowledge base. Overall, NELL demonstrates the potential of never-ending learning, continuously improving its knowledge and capabilities over time.