A key feature of next-generation communications is the ability of radios to not just operate in, but collaborate through, congested spectrum. In recent years, radios have moved well beyond simple cognition like that found in classical approaches to dynamic spectrum access. With the recent multi-disciplinary interest in artificial intelligence and machine learning, radios are becoming more intelligent at all layers of the protocol stack. In fact, we have shown previously through experimental hardware the so-called deepmod: a radio with the ability to design, train, and operate a PHY layer protocol with minimal traditional signal processing blocks. A natural question for deepmod often arises: How does a machine-generated PHY compare with a human-invented PHY which itself has made great strides in adaptability? However, this work attempts to move beyond this question, instead: What behavior must a radio exhibit in order to qualify as intelligent rather than merely resilient or adaptive? Through experimentation with software radios, we explore this, and other, questions related to next-generation machine-assisted communications and radio links.