The primary function of language is to communicate. It is fundamentally about semantics. And semantics represents a major barrier to further progress in language modeling. The sentence gestalt model has very simplistic semantics, and the more advanced version of it developed by Rohde (2002) introduces more complex semantics, at the cost of injecting externally more of what the model should be developing on its own. Thus, the fundamental challenge for models of sentence-level or higher-level language is to develop a more naturalistic way of training the corresponding semantics. In an ideal case, a virtual humanoid robot would be wandering around a rich simulated naturalistic environment, and receiving and producing language in order to understand and survive in this environment. This would mimic the way in which people acquire and use language, and would undoubtedly provide considerable insight into the nature of language acquisition and higher-level semantic representations. But clearly this will require a lot of work.